diff --git a/.flake8 b/.flake8
index 1c0c2cdbb..410cb5482 100644
--- a/.flake8
+++ b/.flake8
@@ -24,6 +24,7 @@ exclude =
   **/data/**,
   icefall/shared/make_kn_lm.py,
   icefall/__init__.py
+  icefall/ctc/__init__.py
 
 ignore =
   # E203 white space before ":"
diff --git a/.github/scripts/run-gigaspeech-zipformer-2023-10-17.sh b/.github/scripts/run-gigaspeech-zipformer-2023-10-17.sh
new file mode 100755
index 000000000..6bb0b9ebc
--- /dev/null
+++ b/.github/scripts/run-gigaspeech-zipformer-2023-10-17.sh
@@ -0,0 +1,94 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/gigaspeech/ASR
+
+repo_url=https://huggingface.co/yfyeung/icefall-asr-gigaspeech-zipformer-2023-10-17
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "data/lang_bpe_500/tokens.txt"
+git lfs pull --include "exp/jit_script.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./zipformer/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./zipformer/jit_pretrained.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --nn-model-filename $repo/exp/jit_script.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for method in greedy_search modified_beam_search fast_beam_search; do
+  log "$method"
+
+  ./zipformer/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p zipformer/exp
+  ln -s $PWD/$repo/exp/pretrained.pt zipformer/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh zipformer/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./zipformer/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir zipformer/exp
+  done
+
+  rm zipformer/exp/*.pt
+fi
diff --git a/.github/scripts/run-pre-trained-conformer-ctc.sh b/.github/scripts/run-pre-trained-conformer-ctc.sh
deleted file mode 100755
index a4959aa01..000000000
--- a/.github/scripts/run-pre-trained-conformer-ctc.sh
+++ /dev/null
@@ -1,46 +0,0 @@
-#!/usr/bin/env bash
-
-set -e
-
-log() {
-  # This function is from espnet
-  local fname=${BASH_SOURCE[1]##*/}
-  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
-}
-
-cd egs/librispeech/ASR
-
-repo_url=https://github.com/csukuangfj/icefall-asr-conformer-ctc-bpe-500
-git lfs install
-
-log "Downloading pre-trained model from $repo_url"
-git clone $repo_url
-repo=$(basename $repo_url)
-
-log "Display test files"
-tree $repo/
-ls -lh $repo/test_wavs/*.flac
-
-log "CTC decoding"
-
-./conformer_ctc/pretrained.py \
-  --method ctc-decoding \
-  --num-classes 500 \
-  --checkpoint $repo/exp/pretrained.pt \
-  --tokens $repo/data/lang_bpe_500/tokens.txt \
-  $repo/test_wavs/1089-134686-0001.flac \
-  $repo/test_wavs/1221-135766-0001.flac \
-  $repo/test_wavs/1221-135766-0002.flac
-
-log "HLG decoding"
-
-./conformer_ctc/pretrained.py \
-  --method 1best \
-  --num-classes 500 \
-  --checkpoint $repo/exp/pretrained.pt \
-  --tokens $repo/data/lang_bpe_500/tokens.txt \
-  --words-file $repo/data/lang_bpe_500/words.txt \
-  --HLG $repo/data/lang_bpe_500/HLG.pt \
-  $repo/test_wavs/1089-134686-0001.flac \
-  $repo/test_wavs/1221-135766-0001.flac \
-  $repo/test_wavs/1221-135766-0002.flac
diff --git a/.github/scripts/run-pre-trained-ctc.sh b/.github/scripts/run-pre-trained-ctc.sh
new file mode 100755
index 000000000..7d6449c9a
--- /dev/null
+++ b/.github/scripts/run-pre-trained-ctc.sh
@@ -0,0 +1,240 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+pushd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/sherpa-onnx-zipformer-ctc-en-2023-10-02
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+log "CTC greedy search"
+
+./zipformer/onnx_pretrained_ctc.py \
+  --nn-model $repo/model.onnx \
+  --tokens $repo/tokens.txt \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+log "CTC H decoding"
+
+./zipformer/onnx_pretrained_ctc_H.py \
+  --nn-model $repo/model.onnx \
+  --tokens $repo/tokens.txt \
+  --H $repo/H.fst \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+log "CTC HL decoding"
+
+./zipformer/onnx_pretrained_ctc_HL.py \
+  --nn-model $repo/model.onnx \
+  --words $repo/words.txt \
+  --HL $repo/HL.fst \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+log "CTC HLG decoding"
+
+./zipformer/onnx_pretrained_ctc_HLG.py \
+  --nn-model $repo/model.onnx \
+  --words $repo/words.txt \
+  --HLG $repo/HLG.fst \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+rm -rf $repo
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+pushd $repo
+
+git lfs pull --include "exp/pretrained.pt"
+git lfs pull --include "data/lang_bpe_500/HLG.pt"
+git lfs pull --include "data/lang_bpe_500/L.pt"
+git lfs pull --include "data/lang_bpe_500/L_disambig.pt"
+git lfs pull --include "data/lang_bpe_500/Linv.pt"
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "data/lang_bpe_500/lexicon.txt"
+git lfs pull --include "data/lang_bpe_500/lexicon_disambig.txt"
+git lfs pull --include "data/lang_bpe_500/tokens.txt"
+git lfs pull --include "data/lang_bpe_500/words.txt"
+git lfs pull --include "data/lm/G_3_gram.fst.txt"
+
+popd
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+log "CTC decoding"
+
+./conformer_ctc/pretrained.py \
+  --method ctc-decoding \
+  --num-classes 500 \
+  --checkpoint $repo/exp/pretrained.pt \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "HLG decoding"
+
+./conformer_ctc/pretrained.py \
+  --method 1best \
+  --num-classes 500 \
+  --checkpoint $repo/exp/pretrained.pt \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --words-file $repo/data/lang_bpe_500/words.txt \
+  --HLG $repo/data/lang_bpe_500/HLG.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "CTC decoding on CPU with kaldi decoders using OpenFst"
+
+log "Exporting model with torchscript"
+
+pushd $repo/exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+./conformer_ctc/export.py \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --jit 1
+
+ls -lh $repo/exp
+
+
+log "Generating H.fst, HL.fst"
+
+./local/prepare_lang_fst.py  --lang-dir $repo/data/lang_bpe_500 --ngram-G $repo/data/lm/G_3_gram.fst.txt
+
+ls -lh $repo/data/lang_bpe_500
+
+log "Decoding with H on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_H.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --H $repo/data/lang_bpe_500/H.fst \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "Decoding with HL on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_HL.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --HL $repo/data/lang_bpe_500/HL.fst \
+  --words $repo/data/lang_bpe_500/words.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "Decoding with HLG on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_HLG.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --HLG $repo/data/lang_bpe_500/HLG.fst \
+  --words $repo/data/lang_bpe_500/words.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+rm -rf $repo
+
+popd
+
+log "Test aishell"
+
+pushd egs/aishell/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall_asr_aishell_conformer_ctc
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+pushd $repo
+
+git lfs pull --include "exp/pretrained.pt"
+git lfs pull --include "data/lang_char/H.fst"
+git lfs pull --include "data/lang_char/HL.fst"
+git lfs pull --include "data/lang_char/HLG.fst"
+
+popd
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+log "CTC decoding"
+
+log "Exporting model with torchscript"
+
+pushd $repo/exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+./conformer_ctc/export.py \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_char/tokens.txt \
+  --jit 1
+
+ls -lh $repo/exp
+
+ls -lh $repo/data/lang_char
+
+log "Decoding with H on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_H.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --H $repo/data/lang_char/H.fst \
+  --tokens $repo/data/lang_char/tokens.txt \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+log "Decoding with HL on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_HL.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --HL $repo/data/lang_char/HL.fst \
+  --words $repo/data/lang_char/words.txt \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+log "Decoding with HLG on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_HLG.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --HLG $repo/data/lang_char/HLG.fst \
+  --words $repo/data/lang_char/words.txt \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+rm -rf $repo
diff --git a/.github/scripts/run-swbd-conformer-ctc-2023-08-26.sh b/.github/scripts/run-swbd-conformer-ctc-2023-08-26.sh
new file mode 100755
index 000000000..d8cc020e1
--- /dev/null
+++ b/.github/scripts/run-swbd-conformer-ctc-2023-08-26.sh
@@ -0,0 +1,44 @@
+#!/usr/bin/env bash
+
+set -e
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/swbd/ASR
+
+repo_url=https://huggingface.co/zrjin/icefall-asr-swbd-conformer-ctc-2023-8-26
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s epoch-98.pt epoch-99.pt
+popd
+
+ls -lh $repo/exp/*.pt
+
+for method in ctc-decoding 1best; do
+  log "$method"
+
+  ./conformer_ctc/pretrained.py \
+    --method $method \
+    --checkpoint $repo/exp/epoch-99.pt \
+    --tokens $repo/data/lang_bpe_500/tokens.txt \
+    --words-file $repo/data/lang_bpe_500/words.txt \
+    --HLG  $repo/data/lang_bpe_500/HLG.pt \
+    --G $repo/data/lm/G_4_gram.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+done
diff --git a/.github/workflows/build-docker-image.yml b/.github/workflows/build-docker-image.yml
index 327f0ee45..e5d96dcdf 100644
--- a/.github/workflows/build-docker-image.yml
+++ b/.github/workflows/build-docker-image.yml
@@ -16,7 +16,7 @@ jobs:
       fail-fast: false
       matrix:
         os: [ubuntu-latest]
-        image: ["torch2.0.0-cuda11.7", "torch1.13.0-cuda11.6", "torch1.12.1-cuda11.3", "torch1.9.0-cuda10.2"]
+        image: ["torch2.1.0-cuda12.1", "torch2.1.0-cuda11.8", "torch2.0.0-cuda11.7", "torch1.13.0-cuda11.6", "torch1.12.1-cuda11.3", "torch1.9.0-cuda10.2"]
 
     steps:
       # refer to https://github.com/actions/checkout
@@ -30,6 +30,13 @@ jobs:
           image=${{ matrix.image }}
           mv -v ./docker/$image.dockerfile ./Dockerfile
 
+      - name: Free space
+        shell: bash
+        run: |
+          df -h
+          rm -rf /opt/hostedtoolcache
+          df -h
+
       - name: Log in to Docker Hub
         uses: docker/login-action@v2
         with:
diff --git a/.github/workflows/run-docker-image.yml b/.github/workflows/run-docker-image.yml
index 12604a132..d048923b6 100644
--- a/.github/workflows/run-docker-image.yml
+++ b/.github/workflows/run-docker-image.yml
@@ -14,7 +14,7 @@ jobs:
       fail-fast: false
       matrix:
         os: [ubuntu-latest]
-        image: ["torch2.0.0-cuda11.7", "torch1.13.0-cuda11.6", "torch1.12.1-cuda11.3", "torch1.9.0-cuda10.2"]
+        image: ["torch2.1.0-cuda12.1", "torch2.1.0-cuda11.8", "torch2.0.0-cuda11.7", "torch1.13.0-cuda11.6", "torch1.12.1-cuda11.3", "torch1.9.0-cuda10.2"]
     steps:
       # refer to https://github.com/actions/checkout
       - uses: actions/checkout@v2
@@ -30,8 +30,15 @@ jobs:
               uname -a
               cat /etc/*release
 
+              find / -name libcuda* 2>/dev/null
+
+              ls -lh /usr/local/
+              ls -lh /usr/local/cuda*
+
               nvcc --version
 
+              ls -lh /usr/local/cuda-*/compat/*
+
               # For torch1.9.0-cuda10.2
               export LD_LIBRARY_PATH=/usr/local/cuda-10.2/compat:$LD_LIBRARY_PATH
 
@@ -41,6 +48,12 @@ jobs:
               # For torch2.0.0-cuda11.7
               export LD_LIBRARY_PATH=/usr/local/cuda-11.7/compat:$LD_LIBRARY_PATH
 
+              # For torch2.1.0-cuda11.8
+              export LD_LIBRARY_PATH=/usr/local/cuda-11.8/compat:$LD_LIBRARY_PATH
+
+              # For torch2.1.0-cuda12.1
+              export LD_LIBRARY_PATH=/usr/local/cuda-12.1/compat:$LD_LIBRARY_PATH
+
 
               which nvcc
               cuda_dir=$(dirname $(which nvcc))
diff --git a/.github/workflows/run-gigaspeech-zipformer-2023-10-17.yml b/.github/workflows/run-gigaspeech-zipformer-2023-10-17.yml
new file mode 100644
index 000000000..7572f4b5f
--- /dev/null
+++ b/.github/workflows/run-gigaspeech-zipformer-2023-10-17.yml
@@ -0,0 +1,126 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-gigaspeech-zipformer-2023-10-17
+# zipformer
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+concurrency:
+  group: run_gigaspeech_2023_10_17_zipformer-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_gigaspeech_2023_10_17_zipformer:
+    if: github.event.label.name == 'zipformer' ||github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/gigaspeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/gigaspeech/ASR/data/fbank
+          ls -lh egs/gigaspeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-gigaspeech-zipformer-2023-10-17.sh
+
+      - name: Display decoding results for gigaspeech zipformer
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/gigaspeech/ASR/
+          tree ./zipformer/exp
+
+          cd zipformer
+          echo "results for zipformer"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for gigaspeech zipformer
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-zipformer-2022-11-11
+          path: egs/gigaspeech/ASR/zipformer/exp/
diff --git a/.github/workflows/run-pretrained-conformer-ctc.yml b/.github/workflows/run-pretrained-ctc.yml
similarity index 86%
rename from .github/workflows/run-pretrained-conformer-ctc.yml
rename to .github/workflows/run-pretrained-ctc.yml
index 6151a5a14..074a63dfc 100644
--- a/.github/workflows/run-pretrained-conformer-ctc.yml
+++ b/.github/workflows/run-pretrained-ctc.yml
@@ -14,7 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-name: run-pre-trained-conformer-ctc
+name: run-pre-trained-ctc
 
 on:
   push:
@@ -23,13 +23,20 @@ on:
   pull_request:
     types: [labeled]
 
+  workflow_dispatch:
+    inputs:
+      test-run:
+        description: 'Test (y/n)?'
+        required: true
+        default: 'y'
+
 concurrency:
-  group: run_pre_trained_conformer_ctc-${{ github.ref }}
+  group: run_pre_trained_ctc-${{ github.ref }}
   cancel-in-progress: true
 
 jobs:
-  run_pre_trained_conformer_ctc:
-    if: github.event.label.name == 'ready' || github.event_name == 'push'
+  run_pre_trained_ctc:
+    if: github.event.label.name == 'ready' || github.event_name == 'push' || github.event.inputs.test-run == 'y' || github.event.label.name == 'ctc'
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
@@ -77,4 +84,4 @@ jobs:
           export PYTHONPATH=$PWD:$PYTHONPATH
           export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
           export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
-          .github/scripts/run-pre-trained-conformer-ctc.sh
+          .github/scripts/run-pre-trained-ctc.sh
diff --git a/.github/workflows/run-swbd-conformer-ctc.yml b/.github/workflows/run-swbd-conformer-ctc.yml
new file mode 100644
index 000000000..842691d38
--- /dev/null
+++ b/.github/workflows/run-swbd-conformer-ctc.yml
@@ -0,0 +1,84 @@
+# Copyright      2023   Xiaomi Corp.    (author: Zengrui Jin)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-swbd-conformer_ctc
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+concurrency:
+  group: run-swbd-conformer_ctc-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run-swbd-conformer_ctc:
+    if: github.event.label.name == 'onnx' || github.event.label.name == 'ready' || github.event_name == 'push' || github.event.label.name == 'swbd'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest]
+        python-version: [3.8]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf==3.20.*
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2023-05-22
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          sudo apt-get -qq install git-lfs tree
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-swbd-conformer-ctc-2023-08-26.sh
diff --git a/.github/workflows/run-yesno-recipe.yml b/.github/workflows/run-yesno-recipe.yml
index 57f15fe87..9ac848535 100644
--- a/.github/workflows/run-yesno-recipe.yml
+++ b/.github/workflows/run-yesno-recipe.yml
@@ -60,12 +60,12 @@ jobs:
 
       - name: Install Python dependencies
         run: |
-          grep -v '^#' ./requirements-ci.txt  | grep -v kaldifst | xargs -n 1 -L 1 pip install
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
           pip uninstall -y protobuf
           pip install --no-binary protobuf protobuf==3.20.*
 
-          pip install --no-deps --force-reinstall https://huggingface.co/csukuangfj/k2/resolve/main/cpu/k2-1.24.3.dev20230508+cpu.torch1.13.1-cp38-cp38-linux_x86_64.whl
-          pip install kaldifeat==1.25.0.dev20230726+cpu.torch1.13.1 -f https://csukuangfj.github.io/kaldifeat/cpu.html
+          pip install --no-deps --force-reinstall k2==1.24.4.dev20231021+cpu.torch1.13.1 -f https://k2-fsa.github.io/k2/cpu.html
+          pip install kaldifeat==1.25.1.dev20231022+cpu.torch1.13.1 -f https://csukuangfj.github.io/kaldifeat/cpu.html
 
       - name: Run yesno recipe
         shell: bash
@@ -140,9 +140,46 @@ jobs:
             download/waves_yesno/0_0_0_1_0_0_0_1.wav \
             download/waves_yesno/0_0_1_0_0_0_1_0.wav
 
+      - name: Test decoding with H
+        shell: bash
+        working-directory: ${{github.workspace}}
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          echo $PYTHONPATH
+
+          cd egs/yesno/ASR
+          python3 ./tdnn/export.py --epoch 14 --avg 2 --jit 1
+
+          python3 ./tdnn/jit_pretrained_decode_with_H.py \
+              --nn-model ./tdnn/exp/cpu_jit.pt \
+              --H ./data/lang_phone/H.fst \
+              --tokens ./data/lang_phone/tokens.txt \
+              ./download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+              ./download/waves_yesno/0_0_1_0_0_0_1_0.wav \
+              ./download/waves_yesno/0_0_1_0_0_1_1_1.wav
+
+      - name: Test decoding with HL
+        shell: bash
+        working-directory: ${{github.workspace}}
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          echo $PYTHONPATH
+
+          cd egs/yesno/ASR
+          python3 ./tdnn/export.py --epoch 14 --avg 2 --jit 1
+
+          python3 ./tdnn/jit_pretrained_decode_with_HL.py \
+              --nn-model ./tdnn/exp/cpu_jit.pt \
+              --HL ./data/lang_phone/HL.fst \
+              --words ./data/lang_phone/words.txt \
+              ./download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+              ./download/waves_yesno/0_0_1_0_0_0_1_0.wav \
+              ./download/waves_yesno/0_0_1_0_0_1_1_1.wav
+
       - name: Show generated files
         shell: bash
         working-directory: ${{github.workspace}}
         run: |
           cd egs/yesno/ASR
           ls -lh tdnn/exp
+          ls -lh data/lang_phone
diff --git a/.gitignore b/.gitignore
index 8af05d884..fa18ca83c 100644
--- a/.gitignore
+++ b/.gitignore
@@ -34,3 +34,5 @@ node_modules
 *.param
 *.bin
 .DS_Store
+*.fst
+*.arpa
diff --git a/README.md b/README.md
index 523203aa4..a14abd023 100644
--- a/README.md
+++ b/README.md
@@ -29,6 +29,7 @@ We provide the following recipes:
   - [yesno][yesno]
   - [LibriSpeech][librispeech]
   - [GigaSpeech][gigaspeech]
+  - [AMI][ami]
   - [Aishell][aishell]
   - [Aishell2][aishell2]
   - [Aishell4][aishell4]
@@ -37,6 +38,7 @@ We provide the following recipes:
   - [Aidatatang_200zh][aidatatang_200zh]
   - [WenetSpeech][wenetspeech]
   - [Alimeeting][alimeeting]
+  - [Switchboard][swbd]
   - [TAL_CSASR][tal_csasr]
 
 ### yesno
@@ -118,9 +120,9 @@ We provide a Colab notebook to run a pre-trained transducer conformer + stateles
 
 | Encoder         | Params | test-clean | test-other |
 |-----------------|--------|------------|------------|
-| zipformer       | 65.5M  | 2.21       | 4.91       |
-| zipformer-small | 23.2M  | 2.46       | 5.83       |
-| zipformer-large | 148.4M | 2.11       | 4.77       |
+| zipformer       | 65.5M  | 2.21       | 4.79       |
+| zipformer-small | 23.2M  | 2.42       | 5.73       |
+| zipformer-large | 148.4M | 2.06       | 4.63       |
 
 Note: No auxiliary losses are used in the training and no LMs are used
 in the decoding.
@@ -146,8 +148,11 @@ in the decoding.
 
 ### GigaSpeech
 
-We provide two models for this recipe: [Conformer CTC model][GigaSpeech_conformer_ctc]
-and [Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss][GigaSpeech_pruned_transducer_stateless2].
+We provide three models for this recipe:
+
+- [Conformer CTC model][GigaSpeech_conformer_ctc]
+- [Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss][GigaSpeech_pruned_transducer_stateless2].
+- [Transducer: Zipformer encoder + Embedding decoder][GigaSpeech_zipformer]
 
 #### Conformer CTC
 
@@ -163,6 +168,14 @@ and [Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned R
 |   fast beam search   | 10.50 | 10.69 |
 | modified beam search | 10.40 | 10.51 |
 
+#### Transducer: Zipformer encoder + Embedding decoder
+
+|                      |  Dev  | Test  |
+|----------------------|-------|-------|
+|    greedy search     | 10.31 | 10.50 |
+|   fast beam search   | 10.26 | 10.48 |
+| modified beam search | 10.25 | 10.38 |
+
 
 ### Aishell
 
@@ -376,6 +389,7 @@ Please see: [![Open In Colab](https://colab.research.google.com/assets/colab-bad
 [TED-LIUM3_pruned_transducer_stateless]: egs/tedlium3/ASR/pruned_transducer_stateless
 [GigaSpeech_conformer_ctc]: egs/gigaspeech/ASR/conformer_ctc
 [GigaSpeech_pruned_transducer_stateless2]: egs/gigaspeech/ASR/pruned_transducer_stateless2
+[GigaSpeech_zipformer]: egs/gigaspeech/ASR/zipformer
 [Aidatatang_200zh_pruned_transducer_stateless2]: egs/aidatatang_200zh/ASR/pruned_transducer_stateless2
 [WenetSpeech_pruned_transducer_stateless2]: egs/wenetspeech/ASR/pruned_transducer_stateless2
 [WenetSpeech_pruned_transducer_stateless5]: egs/wenetspeech/ASR/pruned_transducer_stateless5
@@ -393,4 +407,6 @@ Please see: [![Open In Colab](https://colab.research.google.com/assets/colab-bad
 [wenetspeech]: egs/wenetspeech/ASR
 [alimeeting]: egs/alimeeting/ASR
 [tal_csasr]: egs/tal_csasr/ASR
+[ami]: egs/ami
+[swbd]: egs/swbd/ASR
 [k2]: https://github.com/k2-fsa/k2
diff --git a/docker/torch1.12.1-cuda11.3.dockerfile b/docker/torch1.12.1-cuda11.3.dockerfile
index 5338bdca7..ed746abe3 100644
--- a/docker/torch1.12.1-cuda11.3.dockerfile
+++ b/docker/torch1.12.1-cuda11.3.dockerfile
@@ -4,8 +4,9 @@ ENV LC_ALL C.UTF-8
 
 ARG DEBIAN_FRONTEND=noninteractive
 
-ARG K2_VERSION="1.24.3.dev20230725+cuda11.3.torch1.12.1"
-ARG KALDIFEAT_VERSION="1.25.0.dev20230726+cuda11.3.torch1.12.1"
+# python 3.7
+ARG K2_VERSION="1.24.4.dev20230725+cuda11.3.torch1.12.1"
+ARG KALDIFEAT_VERSION="1.25.1.dev20231022+cuda11.3.torch1.12.1"
 ARG TORCHAUDIO_VERSION="0.12.1+cu113"
 
 LABEL authors="Fangjun Kuang <csukuangfj@gmail.com>"
diff --git a/docker/torch1.13.0-cuda11.6.dockerfile b/docker/torch1.13.0-cuda11.6.dockerfile
index 4d2f96c8e..9657866e5 100644
--- a/docker/torch1.13.0-cuda11.6.dockerfile
+++ b/docker/torch1.13.0-cuda11.6.dockerfile
@@ -4,8 +4,9 @@ ENV LC_ALL C.UTF-8
 
 ARG DEBIAN_FRONTEND=noninteractive
 
-ARG K2_VERSION="1.24.3.dev20230725+cuda11.6.torch1.13.0"
-ARG KALDIFEAT_VERSION="1.25.0.dev20230726+cuda11.6.torch1.13.0"
+# python 3.9
+ARG K2_VERSION="1.24.4.dev20231021+cuda11.6.torch1.13.0"
+ARG KALDIFEAT_VERSION="1.25.1.dev20231022+cuda11.6.torch1.13.0"
 ARG TORCHAUDIO_VERSION="0.13.0+cu116"
 
 LABEL authors="Fangjun Kuang <csukuangfj@gmail.com>"
diff --git a/docker/torch1.9.0-cuda10.2.dockerfile b/docker/torch1.9.0-cuda10.2.dockerfile
index a7cef6dc8..a92af7ad0 100644
--- a/docker/torch1.9.0-cuda10.2.dockerfile
+++ b/docker/torch1.9.0-cuda10.2.dockerfile
@@ -4,8 +4,9 @@ ENV LC_ALL C.UTF-8
 
 ARG DEBIAN_FRONTEND=noninteractive
 
+# python 3.7
 ARG K2_VERSION="1.24.3.dev20230726+cuda10.2.torch1.9.0"
-ARG KALDIFEAT_VERSION="1.25.0.dev20230726+cuda10.2.torch1.9.0"
+ARG KALDIFEAT_VERSION="1.25.1.dev20231022+cuda10.2.torch1.9.0"
 ARG TORCHAUDIO_VERSION="0.9.0"
 
 LABEL authors="Fangjun Kuang <csukuangfj@gmail.com>"
diff --git a/docker/torch2.0.0-cuda11.7.dockerfile b/docker/torch2.0.0-cuda11.7.dockerfile
index d91fbc24f..07296e6f0 100644
--- a/docker/torch2.0.0-cuda11.7.dockerfile
+++ b/docker/torch2.0.0-cuda11.7.dockerfile
@@ -4,8 +4,9 @@ ENV LC_ALL C.UTF-8
 
 ARG DEBIAN_FRONTEND=noninteractive
 
-ARG K2_VERSION="1.24.3.dev20230718+cuda11.7.torch2.0.0"
-ARG KALDIFEAT_VERSION="1.25.0.dev20230726+cuda11.7.torch2.0.0"
+# python 3.10
+ARG K2_VERSION="1.24.4.dev20231021+cuda11.7.torch2.0.0"
+ARG KALDIFEAT_VERSION="1.25.1.dev20231022+cuda11.7.torch2.0.0"
 ARG TORCHAUDIO_VERSION="2.0.0+cu117"
 
 LABEL authors="Fangjun Kuang <csukuangfj@gmail.com>"
diff --git a/docker/torch2.1.0-cuda11.8.dockerfile b/docker/torch2.1.0-cuda11.8.dockerfile
new file mode 100644
index 000000000..e500e9a6a
--- /dev/null
+++ b/docker/torch2.1.0-cuda11.8.dockerfile
@@ -0,0 +1,71 @@
+FROM pytorch/pytorch:2.1.0-cuda11.8-cudnn8-devel
+
+ENV LC_ALL C.UTF-8
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# python 3.10
+ARG K2_VERSION="1.24.4.dev20231021+cuda11.8.torch2.1.0"
+ARG KALDIFEAT_VERSION="1.25.1.dev20231022+cuda11.8.torch2.1.0"
+ARG TORCHAUDIO_VERSION="2.1.0+cu118"
+
+LABEL authors="Fangjun Kuang <csukuangfj@gmail.com>"
+LABEL k2_version=${K2_VERSION}
+LABEL kaldifeat_version=${KALDIFEAT_VERSION}
+LABEL github_repo="https://github.com/k2-fsa/icefall"
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        curl \
+        vim \
+    		libssl-dev \
+        autoconf \
+        automake \
+        bzip2 \
+        ca-certificates \
+        ffmpeg \
+        g++ \
+        gfortran \
+        git \
+        libtool \
+        make \
+        patch \
+        sox \
+        subversion \
+        unzip \
+        valgrind \
+        wget \
+        zlib1g-dev \
+        && rm -rf /var/lib/apt/lists/*
+
+# Install dependencies
+RUN pip install --no-cache-dir \
+      torchaudio==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torch_stable.html \
+      k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \
+      git+https://github.com/lhotse-speech/lhotse \
+      kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \
+      \
+      kaldi_native_io \
+      kaldialign \
+      kaldifst \
+      kaldilm \
+      sentencepiece>=0.1.96 \
+      tensorboard \
+      typeguard \
+      dill \
+      onnx \
+      onnxruntime \
+      onnxmltools \
+      multi_quantization \
+      typeguard \
+      numpy \
+      pytest \
+      graphviz
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+    cd /workspace/icefall && \
+    pip install --no-cache-dir -r requirements.txt
+
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+WORKDIR /workspace/icefall
diff --git a/docker/torch2.1.0-cuda12.1.dockerfile b/docker/torch2.1.0-cuda12.1.dockerfile
new file mode 100644
index 000000000..c3f12323e
--- /dev/null
+++ b/docker/torch2.1.0-cuda12.1.dockerfile
@@ -0,0 +1,71 @@
+FROM pytorch/pytorch:2.1.0-cuda12.1-cudnn8-devel
+
+ENV LC_ALL C.UTF-8
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# python 3.10
+ARG K2_VERSION="1.24.4.dev20231021+cuda12.1.torch2.1.0"
+ARG KALDIFEAT_VERSION="1.25.1.dev20231022+cuda12.1.torch2.1.0"
+ARG TORCHAUDIO_VERSION="2.1.0+cu121"
+
+LABEL authors="Fangjun Kuang <csukuangfj@gmail.com>"
+LABEL k2_version=${K2_VERSION}
+LABEL kaldifeat_version=${KALDIFEAT_VERSION}
+LABEL github_repo="https://github.com/k2-fsa/icefall"
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        curl \
+        vim \
+    		libssl-dev \
+        autoconf \
+        automake \
+        bzip2 \
+        ca-certificates \
+        ffmpeg \
+        g++ \
+        gfortran \
+        git \
+        libtool \
+        make \
+        patch \
+        sox \
+        subversion \
+        unzip \
+        valgrind \
+        wget \
+        zlib1g-dev \
+        && rm -rf /var/lib/apt/lists/*
+
+# Install dependencies
+RUN pip install --no-cache-dir \
+      torchaudio==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torch_stable.html \
+      k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \
+      git+https://github.com/lhotse-speech/lhotse \
+      kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \
+      \
+      kaldi_native_io \
+      kaldialign \
+      kaldifst \
+      kaldilm \
+      sentencepiece>=0.1.96 \
+      tensorboard \
+      typeguard \
+      dill \
+      onnx \
+      onnxruntime \
+      onnxmltools \
+      multi_quantization \
+      typeguard \
+      numpy \
+      pytest \
+      graphviz
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+    cd /workspace/icefall && \
+    pip install --no-cache-dir -r requirements.txt
+
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+WORKDIR /workspace/icefall
diff --git a/docs/source/contributing/code-style.rst b/docs/source/contributing/code-style.rst
index 3baaaeec2..cb08229c3 100644
--- a/docs/source/contributing/code-style.rst
+++ b/docs/source/contributing/code-style.rst
@@ -38,7 +38,7 @@ Please fix any issues reported by the check tools.
 .. HINT::
 
   Some of the check tools, i.e., ``black`` and ``isort`` will modify
-  the files to be commited **in-place**. So please run ``git status``
+  the files to be committed **in-place**. So please run ``git status``
   after failure to see which file has been modified by the tools
   before you make any further changes.
 
diff --git a/docs/source/decoding-with-langugage-models/LODR.rst b/docs/source/decoding-with-langugage-models/LODR.rst
index 8cc1a624c..b6b6e8cbb 100644
--- a/docs/source/decoding-with-langugage-models/LODR.rst
+++ b/docs/source/decoding-with-langugage-models/LODR.rst
@@ -56,7 +56,7 @@ during decoding for transducer model:
     \lambda_1 \log p_{\text{Target LM}}\left(y_u|\mathit{x},y_{1:u-1}\right) -
     \lambda_2 \log p_{\text{bi-gram}}\left(y_u|\mathit{x},y_{1:u-1}\right)
 
-In LODR, an additional bi-gram LM estimated on the source domain (e.g training corpus) is required. Comared to DR,
+In LODR, an additional bi-gram LM estimated on the source domain (e.g training corpus) is required. Compared to DR,
 the only difference lies in the choice of source domain LM. According to the original `paper <https://arxiv.org/abs/2203.16776>`_,
 LODR achieves similar performance compared DR in both intra-domain and cross-domain settings.
 As a bi-gram is much faster to evaluate, LODR is usually much faster.
diff --git a/docs/source/decoding-with-langugage-models/index.rst b/docs/source/decoding-with-langugage-models/index.rst
index 6e5e3a4d9..c49da9a4e 100644
--- a/docs/source/decoding-with-langugage-models/index.rst
+++ b/docs/source/decoding-with-langugage-models/index.rst
@@ -2,12 +2,13 @@ Decoding with language models
 =============================
 
 This section describes how to use external langugage models 
-during decoding to improve the WER of transducer models.
+during decoding to improve the WER of transducer models. To train an external language model,
+please refer to this tutorial: :ref:`train_nnlm`.
 
 The following decoding methods with external langugage models are available:
 
 
-.. list-table:: LM-rescoring-based methods vs shallow-fusion-based methods (The numbers in each field is WER on test-clean, WER on test-other and decoding time on test-clean)
+.. list-table:: 
    :widths: 25 50
    :header-rows: 1
 
diff --git a/docs/source/docker/intro.rst b/docs/source/docker/intro.rst
index b09247d85..9ead0df00 100644
--- a/docs/source/docker/intro.rst
+++ b/docs/source/docker/intro.rst
@@ -30,6 +30,8 @@ which will give you something like below:
 
 .. code-block:: bash
 
+  "torch2.1.0-cuda12.1"
+  "torch2.1.0-cuda11.8"
   "torch2.0.0-cuda11.7"
   "torch1.12.1-cuda11.3"
   "torch1.9.0-cuda10.2"
diff --git a/docs/source/model-export/export-ncnn-conv-emformer.rst b/docs/source/model-export/export-ncnn-conv-emformer.rst
index 4f5535d83..93392aee7 100644
--- a/docs/source/model-export/export-ncnn-conv-emformer.rst
+++ b/docs/source/model-export/export-ncnn-conv-emformer.rst
@@ -125,7 +125,7 @@ Python code. We have also set up ``PATH`` so that you can use
 .. caution::
 
   Please don't use `<https://github.com/tencent/ncnn>`_.
-  We have made some modifications to the offical `ncnn`_.
+  We have made some modifications to the official `ncnn`_.
 
   We will synchronize `<https://github.com/csukuangfj/ncnn>`_ periodically
   with the official one.
diff --git a/docs/source/model-export/export-ncnn.rst b/docs/source/model-export/export-ncnn.rst
index 9eb5f85d2..634fb1e59 100644
--- a/docs/source/model-export/export-ncnn.rst
+++ b/docs/source/model-export/export-ncnn.rst
@@ -1,3 +1,5 @@
+.. _icefall_export_to_ncnn:
+
 Export to ncnn
 ==============
 
diff --git a/docs/source/recipes/Non-streaming-ASR/librispeech/distillation.rst b/docs/source/recipes/Non-streaming-ASR/librispeech/distillation.rst
index 2e8d0893a..37edf7de9 100644
--- a/docs/source/recipes/Non-streaming-ASR/librispeech/distillation.rst
+++ b/docs/source/recipes/Non-streaming-ASR/librispeech/distillation.rst
@@ -47,7 +47,7 @@ The data preparation contains several stages, you can use the following two
 options:
 
   - ``--stage``
-  - ``--stop-stage``
+  - ``--stop_stage``
 
 to control which stage(s) should be run. By default, all stages are executed.
 
@@ -56,8 +56,8 @@ For example,
 .. code-block:: bash
 
   $ cd egs/librispeech/ASR
-  $ ./prepare.sh --stage 0 --stop-stage 0 # run only stage 0
-  $ ./prepare.sh --stage 2 --stop-stage 5 # run from stage 2 to stage 5
+  $ ./prepare.sh --stage 0 --stop_stage 0 # run only stage 0
+  $ ./prepare.sh --stage 2 --stop_stage 5 # run from stage 2 to stage 5
 
 .. HINT::
 
@@ -108,15 +108,15 @@ As usual, you can control the stages you want to run by specifying the following
 two options:
 
   - ``--stage``
-  - ``--stop-stage``
+  - ``--stop_stage``
 
 For example,
 
 .. code-block:: bash
 
   $ cd egs/librispeech/ASR
-  $ ./distillation_with_hubert.sh --stage 0 --stop-stage 0 # run only stage 0
-  $ ./distillation_with_hubert.sh --stage 2 --stop-stage 4 # run from stage 2 to stage 5
+  $ ./distillation_with_hubert.sh --stage 0 --stop_stage 0 # run only stage 0
+  $ ./distillation_with_hubert.sh --stage 2 --stop_stage 4 # run from stage 2 to stage 5
 
 Here are a few options in `./distillation_with_hubert.sh <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/distillation_with_hubert.sh>`_
 you need to know before you proceed.
@@ -134,7 +134,7 @@ and prepares MVQ-augmented training manifests.
 
 .. code-block:: bash
 
-  $ ./distillation_with_hubert.sh --stage 2 --stop-stage 2 # run only stage 2
+  $ ./distillation_with_hubert.sh --stage 2 --stop_stage 2 # run only stage 2
 
 Please see the
 following screenshot for the output of an example execution.
@@ -172,7 +172,7 @@ To perform training, please run stage 3 by executing the following command.
 
 .. code-block:: bash
 
-  $ ./prepare.sh --stage 3 --stop-stage 3 # run MVQ training
+  $ ./prepare.sh --stage 3 --stop_stage 3 # run MVQ training
 
 Here is the code snippet for training:
 
diff --git a/docs/source/recipes/RNN-LM/index.rst b/docs/source/recipes/RNN-LM/index.rst
new file mode 100644
index 000000000..4b74e64c7
--- /dev/null
+++ b/docs/source/recipes/RNN-LM/index.rst
@@ -0,0 +1,7 @@
+RNN-LM
+======
+
+.. toctree::
+   :maxdepth: 2
+
+   librispeech/lm-training
\ No newline at end of file
diff --git a/docs/source/recipes/RNN-LM/librispeech/lm-training.rst b/docs/source/recipes/RNN-LM/librispeech/lm-training.rst
new file mode 100644
index 000000000..736120275
--- /dev/null
+++ b/docs/source/recipes/RNN-LM/librispeech/lm-training.rst
@@ -0,0 +1,104 @@
+.. _train_nnlm:
+
+Train an RNN langugage model
+======================================
+
+If you have enough text data, you can train a neural network language model (NNLM) to improve
+the WER of your E2E ASR system. This tutorial shows you how to train an RNNLM from 
+scratch.
+
+.. HINT::
+
+    For how to use an NNLM during decoding, please refer to the following tutorials:
+    :ref:`shallow_fusion`, :ref:`LODR`, :ref:`rescoring`
+
+.. note::
+
+    This tutorial is based on the LibriSpeech recipe. Please check it out for the necessary
+    python scripts for this tutorial. We use the LibriSpeech LM-corpus as the LM training set 
+    for illustration purpose. You can also collect your own data. The data format is quite simple:
+    each line should contain a complete sentence, and words should be separated by space.
+
+First, let's download the training data for the RNNLM. This can be done via the 
+following command:
+
+.. code-block:: bash
+
+    $ wget https://www.openslr.org/resources/11/librispeech-lm-norm.txt.gz 
+    $ gzip -d librispeech-lm-norm.txt.gz
+
+As we are training a BPE-level RNNLM, we need to tokenize the training text, which requires a
+BPE tokenizer. This can be achieved by executing the following command:
+
+.. code-block:: bash
+    
+    $ # if you don't have the BPE
+    $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15
+    $ cd icefall-asr-librispeech-zipformer-2023-05-15/data/lang_bpe_500
+    $ git lfs pull --include bpe.model
+    $ cd ../../..
+
+    $ ./local/prepare_lm_training_data.py \
+        --bpe-model icefall-asr-librispeech-zipformer-2023-05-15/data/lang_bpe_500/bpe.model \
+        --lm-data librispeech-lm-norm.txt \
+        --lm-archive data/lang_bpe_500/lm_data.pt
+
+Now, you should have a file name ``lm_data.pt`` file store under the directory ``data/lang_bpe_500``.
+This is the packed training data for the RNNLM. We then sort the training data according to its
+sentence length.
+
+.. code-block:: bash
+
+    $ # This could take a while (~ 20 minutes), feel free to grab a cup of coffee :)
+    $ ./local/sort_lm_training_data.py \
+        --in-lm-data data/lang_bpe_500/lm_data.pt \
+        --out-lm-data data/lang_bpe_500/sorted_lm_data.pt \
+        --out-statistics data/lang_bpe_500/lm_data_stats.txt
+
+
+The aforementioned steps can be repeated to create a a validation set for you RNNLM. Let's say 
+you have a validation set in ``valid.txt``, you can just set ``--lm-data valid.txt`` 
+and ``--lm-archive data/lang_bpe_500/lm-data-valid.pt`` when calling ``./local/prepare_lm_training_data.py``.
+
+After completing the previous steps, the training and testing sets for training RNNLM are ready. 
+The next step is to train the RNNLM model. The training command is as follows:
+
+.. code-block:: bash
+
+    $ # assume you are in the icefall root directory
+    $ cd rnn_lm
+    $ ln -s ../../egs/librispeech/ASR/data .
+    $ cd ..
+    $ ./rnn_lm/train.py \
+        --world-size 4 \
+        --exp-dir ./rnn_lm/exp \
+        --start-epoch 0 \
+        --num-epochs 10 \
+        --use-fp16 0 \
+        --tie-weights 1 \
+        --embedding-dim 2048 \
+        --hidden_dim 2048 \
+        --num-layers 3 \
+        --batch-size 300 \
+        --lm-data rnn_lm/data/lang_bpe_500/sorted_lm_data.pt \
+        --lm-data-valid rnn_lm/data/lang_bpe_500/sorted_lm_data.pt
+
+
+.. note::
+
+    You can adjust the RNNLM hyper parameters to control the size of the RNNLM,
+    such as embedding dimension and hidden state dimension. For more details, please
+    run ``./rnn_lm/train.py --help``.
+
+.. note::
+
+    The training of RNNLM can take a long time (usually a couple of days).
+
+
+
+
+
+
+
+
+
diff --git a/docs/source/recipes/index.rst b/docs/source/recipes/index.rst
index 63793275c..7265e1cf6 100644
--- a/docs/source/recipes/index.rst
+++ b/docs/source/recipes/index.rst
@@ -15,3 +15,4 @@ We may add recipes for other tasks as well in the future.
 
    Non-streaming-ASR/index
    Streaming-ASR/index
+   RNN-LM/index
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/asr_datamodule.py
index 8f57a303b..d491996b2 100644
--- a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/asr_datamodule.py
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/asr_datamodule.py
@@ -211,7 +211,7 @@ class Aidatatang_200zhAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/train.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/train.py
index c9d9c4aa8..fa809b768 100644
--- a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/train.py
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/train.py
@@ -635,7 +635,6 @@ def train_one_epoch(
     tot_loss = MetricsTracker()
 
     for batch_idx, batch in enumerate(train_dl):
-
         params.batch_idx_train += 1
         batch_size = len(batch["supervisions"]["text"])
 
@@ -800,7 +799,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/aishell/ASR/RESULTS.md b/egs/aishell/ASR/RESULTS.md
index 79026112e..c2b602b0b 100644
--- a/egs/aishell/ASR/RESULTS.md
+++ b/egs/aishell/ASR/RESULTS.md
@@ -145,6 +145,56 @@ for m in greedy_search modified_beam_search fast_beam_search ; do
 done
 ```
 
+#### Pruned transducer stateless 7 streaming
+[./pruned_transducer_stateless7_streaming](./pruned_transducer_stateless7_streaming)
+
+It's Streaming version of Zipformer1 with Pruned RNNT loss.
+
+|                        | test | dev  | comment                               |
+|------------------------|------|------|---------------------------------------|
+| greedy search          | 6.95 | 6.29 | --epoch 44 --avg 15 --max-duration 600 |
+| modified beam search   | 6.51 | 5.90 | --epoch 44 --avg 15 --max-duration 600 |
+| fast beam search       | 6.73 | 6.09 | --epoch 44 --avg 15 --max-duration 600 |
+
+Training command is:
+
+```bash
+./prepare.sh
+
+export CUDA_VISIBLE_DEVICES="0,1"
+
+./pruned_transducer_stateless7_streaming/train.py \
+    --world-size 2 \
+    --num-epochs 50 \
+    --use-fp16 1 \
+    --context-size 1 \
+    --max-duration 800 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --enable-musan 0 \
+    --spec-aug-time-warp-factor 20
+```
+
+**Caution**: It uses `--context-size=1`.
+
+The decoding command is:
+```bash
+for m in greedy_search modified_beam_search fast_beam_search ; do
+  ./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 44 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --lang-dir data/lang_char \
+    --context-size 1 \
+    --decoding-method $m
+done
+```
+
+Pretrained models, training logs, decoding logs, tensorboard and decoding results
+are available at
+<https://huggingface.co/zrjin/icefall-asr-aishell-zipformer-pruned-transducer-stateless7-streaming-2023-10-16/>
+
+
+
 #### Pruned transducer stateless 7
 
 [./pruned_transducer_stateless7](./pruned_transducer_stateless7)
diff --git a/egs/aishell/ASR/conformer_ctc/export.py b/egs/aishell/ASR/conformer_ctc/export.py
old mode 100644
new mode 100755
index 1df3cfdc2..49871d437
--- a/egs/aishell/ASR/conformer_ctc/export.py
+++ b/egs/aishell/ASR/conformer_ctc/export.py
@@ -23,12 +23,12 @@ import argparse
 import logging
 from pathlib import Path
 
+import k2
 import torch
 from conformer import Conformer
 
 from icefall.checkpoint import average_checkpoints, load_checkpoint
-from icefall.lexicon import Lexicon
-from icefall.utils import AttributeDict, str2bool
+from icefall.utils import AttributeDict, num_tokens, str2bool
 
 
 def get_parser():
@@ -63,11 +63,10 @@ def get_parser():
     )
 
     parser.add_argument(
-        "--lang-dir",
+        "--tokens",
         type=str,
-        default="data/lang_char",
-        help="""It contains language related input files such as "lexicon.txt"
-        """,
+        required=True,
+        help="Path to the tokens.txt.",
     )
 
     parser.add_argument(
@@ -98,16 +97,16 @@ def get_params() -> AttributeDict:
 def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
-    args.lang_dir = Path(args.lang_dir)
 
     params = get_params()
     params.update(vars(args))
 
-    logging.info(params)
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
 
-    lexicon = Lexicon(params.lang_dir)
-    max_token_id = max(lexicon.tokens)
-    num_classes = max_token_id + 1  # +1 for the blank
+    num_classes = num_tokens(token_table) + 1  # +1 for the blank
+
+    logging.info(params)
 
     device = torch.device("cpu")
     if torch.cuda.is_available():
diff --git a/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_H.py b/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_H.py
new file mode 120000
index 000000000..896b78aef
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_H.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_H.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py b/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py
new file mode 120000
index 000000000..aa1b6073d
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py b/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py
new file mode 120000
index 000000000..0cf42ce30
--- /dev/null
+++ b/egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/conformer_ctc/test_transformer.py b/egs/aishell/ASR/conformer_ctc/test_transformer.py
old mode 100644
new mode 100755
diff --git a/egs/aishell/ASR/local/prepare_lang_fst.py b/egs/aishell/ASR/local/prepare_lang_fst.py
new file mode 120000
index 000000000..c5787c534
--- /dev/null
+++ b/egs/aishell/ASR/local/prepare_lang_fst.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_fst.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/local/train_bbpe_model.py b/egs/aishell/ASR/local/train_bbpe_model.py
index d231d5d77..48160897d 100755
--- a/egs/aishell/ASR/local/train_bbpe_model.py
+++ b/egs/aishell/ASR/local/train_bbpe_model.py
@@ -15,7 +15,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-
 # You can install sentencepiece via:
 #
 #  pip install sentencepiece
@@ -26,12 +25,12 @@
 # Please install a version >=0.1.96
 
 import argparse
-import re
 import shutil
 import tempfile
 from pathlib import Path
 
 import sentencepiece as spm
+
 from icefall import byte_encode, tokenize_by_CJK_char
 
 
@@ -74,6 +73,11 @@ def main():
     model_type = "unigram"
 
     model_prefix = f"{lang_dir}/{model_type}_{vocab_size}"
+    model_file = Path(model_prefix + ".model")
+    if model_file.is_file():
+        print(f"{model_file} exists - skipping")
+        return
+
     character_coverage = 1.0
     input_sentence_size = 100000000
 
@@ -88,23 +92,18 @@ def main():
 
     _convert_to_bchar(args.transcript, train_text)
 
-    model_file = Path(model_prefix + ".model")
-    if not model_file.is_file():
-        spm.SentencePieceTrainer.train(
-            input=train_text,
-            vocab_size=vocab_size,
-            model_type=model_type,
-            model_prefix=model_prefix,
-            input_sentence_size=input_sentence_size,
-            character_coverage=character_coverage,
-            user_defined_symbols=user_defined_symbols,
-            unk_id=unk_id,
-            bos_id=-1,
-            eos_id=-1,
-        )
-    else:
-        print(f"{model_file} exists - skipping")
-        return
+    spm.SentencePieceTrainer.train(
+        input=train_text,
+        vocab_size=vocab_size,
+        model_type=model_type,
+        model_prefix=model_prefix,
+        input_sentence_size=input_sentence_size,
+        character_coverage=character_coverage,
+        user_defined_symbols=user_defined_symbols,
+        unk_id=unk_id,
+        bos_id=-1,
+        eos_id=-1,
+    )
 
     shutil.copyfile(model_file, f"{lang_dir}/bbpe.model")
 
diff --git a/egs/aishell/ASR/prepare.sh b/egs/aishell/ASR/prepare.sh
index efeffe480..4feed55a8 100755
--- a/egs/aishell/ASR/prepare.sh
+++ b/egs/aishell/ASR/prepare.sh
@@ -144,6 +144,7 @@ if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
     ./local/prepare_lang.py --lang-dir $lang_phone_dir
   fi
 
+
   # Train a bigram P for MMI training
   if [ ! -f $lang_phone_dir/transcript_words.txt ]; then
     log "Generate data to train phone based bigram P"
@@ -258,6 +259,13 @@ if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
       --max-order=3 \
       data/lm/3-gram.unpruned.arpa > data/lm/G_3_gram_char.fst.txt
   fi
+
+  if [ ! -f $lang_char_dir/HLG.fst ]; then
+    lang_phone_dir=data/lang_phone
+    ./local/prepare_lang_fst.py  \
+      --lang-dir $lang_phone_dir \
+      --ngram-G ./data/lm/G_3_gram.fst.txt
+  fi
 fi
 
 if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
diff --git a/egs/aishell/ASR/pruned_transducer_stateless2/train.py b/egs/aishell/ASR/pruned_transducer_stateless2/train.py
index d08908238..60f014c48 100755
--- a/egs/aishell/ASR/pruned_transducer_stateless2/train.py
+++ b/egs/aishell/ASR/pruned_transducer_stateless2/train.py
@@ -872,7 +872,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/train.py b/egs/aishell/ASR/pruned_transducer_stateless3/train.py
index 62e67530d..7c23041ca 100755
--- a/egs/aishell/ASR/pruned_transducer_stateless3/train.py
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/train.py
@@ -1045,7 +1045,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/train.py b/egs/aishell/ASR/pruned_transducer_stateless7/train.py
index cbb7db086..9d9dd4288 100755
--- a/egs/aishell/ASR/pruned_transducer_stateless7/train.py
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/train.py
@@ -703,7 +703,7 @@ def compute_loss(
             if batch_idx_train >= warm_step
             else 0.1 + 0.9 * (batch_idx_train / warm_step)
         )
-        loss = params.simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
 
     assert loss.requires_grad == is_training
 
@@ -1028,7 +1028,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/train2.py b/egs/aishell/ASR/pruned_transducer_stateless7/train2.py
index c30f6f960..057af297f 100755
--- a/egs/aishell/ASR/pruned_transducer_stateless7/train2.py
+++ b/egs/aishell/ASR/pruned_transducer_stateless7/train2.py
@@ -1031,7 +1031,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/train.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/train.py
index 4e52f9573..3858bafd7 100755
--- a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/train.py
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/train.py
@@ -1019,7 +1019,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/README.md b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/README.md
new file mode 120000
index 000000000..a784292cd
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/README.md
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/README.md
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py
new file mode 120000
index 000000000..a074d6085
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/asr_datamodule.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/beam_search.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/beam_search.py
new file mode 120000
index 000000000..8554e44cc
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/beam_search.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py
new file mode 100755
index 000000000..f5ae836fd
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py
@@ -0,0 +1,735 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless7_streaming/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --max-duration 600 \
+    --decode-chunk-len 32 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+import os
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall import ContextGraph
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--context-score",
+        type=float,
+        default=2,
+        help="""
+        The bonus score of each token for the context biasing words/phrases.
+        Used only when --decoding_method is modified_beam_search.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-file",
+        type=str,
+        default="",
+        help="""
+        The path of the context biasing lists, one word/phrase each line
+        Used only when --decoding_method is modified_beam_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps token ID to a string.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    feature_lens += 30
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, 30),
+        value=LOG_EPS,
+    )
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            context_graph=context_graph,
+        )
+    else:
+        hyp_tokens = []
+        batch_size = encoder_out.size(0)
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyp_tokens.append(hyp)
+
+    hyps = [[token_table[t] for t in tokens] for tokens in hyp_tokens]
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        key = f"beam_size_{params.beam_size}"
+        if params.has_contexts:
+            key += f"-context-score-{params.context_score}"
+        else:
+            key += "-no-context-words"
+        return {key: hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps a token ID to a string.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            token_table=token_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+
+    if os.path.exists(params.context_file):
+        params.has_contexts = True
+    else:
+        params.has_contexts = False
+
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        if params.has_contexts:
+            params.suffix += f"-context-score-{params.context_score}"
+        else:
+            params.suffix += "-no-contexts-words"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, (
+        model.encoder.decode_chunk_size,
+        params.decode_chunk_len,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    if params.decoding_method == "modified_beam_search":
+        if os.path.exists(params.context_file):
+            contexts_text = []
+            for line in open(params.context_file).readlines():
+                contexts_text.append(line.strip())
+            contexts = graph_compiler.texts_to_ids(contexts_text)
+            context_graph = ContextGraph(params.context_score)
+            context_graph.build(contexts)
+        else:
+            context_graph = None
+    else:
+        context_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    aishell = AishellAsrDataModule(args)
+
+    test_cuts = aishell.test_cuts()
+    dev_cuts = aishell.valid_cuts()
+
+    test_dl = aishell.test_dataloaders(test_cuts)
+    dev_dl = aishell.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_dls = [test_dl, dev_dl]
+
+    import time
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        start = time.time()
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+        )
+        logging.info(f"Elasped time for {test_set}: {time.time() - start}")
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode_stream.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
new file mode 120000
index 000000000..ca8fed319
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/decode_stream.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decoder.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decoder.py
new file mode 120000
index 000000000..33944d0d2
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decoder.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/decoder.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
new file mode 120000
index 000000000..b9aa0ae08
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/encoder_interface.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
new file mode 120000
index 000000000..72e43c297
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
new file mode 120000
index 000000000..3b36924ef
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py
new file mode 120000
index 000000000..eca5e2956
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx-zh.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx.py
new file mode 120000
index 000000000..57a0cd0a0
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export-onnx.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export-onnx.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export.py
new file mode 120000
index 000000000..2acafdc61
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/export.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py
new file mode 120000
index 000000000..5d9c6ba00
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
new file mode 120000
index 000000000..457131699
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_export.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
new file mode 120000
index 000000000..2b8fa3cbb
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/jit_trace_pretrained.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/joiner.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/joiner.py
new file mode 120000
index 000000000..ecfb6dd8a
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/joiner.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/joiner.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/model.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/model.py
new file mode 120000
index 000000000..e17d4f734
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/model.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/model.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py
new file mode 120000
index 000000000..8eea90e04
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_check.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_check.py
new file mode 120000
index 000000000..28bf7bb82
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_check.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_check.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
new file mode 120000
index 000000000..c8548d459
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_model_wrapper.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py
new file mode 120000
index 000000000..ae4d9bb04
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/optim.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/optim.py
new file mode 120000
index 000000000..81ac4a89a
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/optim.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/optim.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/pretrained.py
new file mode 120000
index 000000000..9510b8fde
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/pretrained.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling.py
new file mode 120000
index 000000000..2428b74b9
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
new file mode 120000
index 000000000..b8b8ba432
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/scaling_converter.py
@@ -0,0 +1 @@
+../pruned_transducer_stateless7/scaling_converter.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
new file mode 120000
index 000000000..92c3904af
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
new file mode 120000
index 000000000..1199a61d6
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
new file mode 100755
index 000000000..6b4f183cf
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py
@@ -0,0 +1,627 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./pruned_transducer_stateless7_streaming/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --decode-chunk-len 32 \
+  --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+  --decoding-method greedy_search \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+import os
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import torch
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import stack_states, unstack_states
+
+from icefall import ContextGraph
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(params.decode_chunk_len)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # We subsample features with ((x_len - 7) // 2 + 1) // 2 and the max downsampling
+    # factor in encoders is 8.
+    # After feature embedding (x_len - 7) // 2, we have (23 - 7) // 2 = 8.
+    tail_length = 23
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+    processed_lens = torch.tensor(processed_lens, device=device)
+
+    encoder_out, encoder_out_lens, new_states = model.encoder.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        states=states,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 50
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = model.encoder.get_init_state(device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=params.decode_chunk_len)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        [
+                            token_table[result]
+                            for result in decode_streams[i].decoding_result()
+                        ],
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    [
+                        token_table[result]
+                        for result in decode_streams[i].decoding_result()
+                    ],
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    # for streaming
+    params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+    )
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    if params.decoding_method == "modified_beam_search":
+        if os.path.exists(params.context_file):
+            contexts_text = []
+            for line in open(params.context_file).readlines():
+                contexts_text.append(line.strip())
+            contexts = graph_compiler.texts_to_ids(contexts_text)
+            context_graph = ContextGraph(params.context_score)
+            context_graph.build(contexts)
+        else:
+            context_graph = None
+    else:
+        context_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    aishell = AishellAsrDataModule(args)
+
+    test_cuts = aishell.test_cuts()
+    valid_cuts = aishell.valid_cuts()
+
+    test_sets = ["test", "valid"]
+    cuts = [test_cuts, valid_cuts]
+
+    for test_set, test_cut in zip(test_sets, cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/test_model.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/test_model.py
new file mode 120000
index 000000000..1259849e0
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/test_model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/test_model.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train.py
new file mode 100755
index 000000000..2e1044658
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train.py
@@ -0,0 +1,1251 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    aishell = AishellAsrDataModule(args)
+
+    train_cuts = aishell.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+            # In pruned RNN-T, we require that T >= S
+            # where T is the number of feature frames after subsampling
+            # and S is the number of tokens in the utterance
+
+            # In ./zipformer.py, the conv module uses the following expression
+            # for subsampling
+            # T = ((c.num_frames - 7) // 2 + 1) // 2
+            # tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+            # if T < len(tokens):
+            #     logging.warning(
+            #         f"Exclude cut with ID {c.id} from training. "
+            #         f"Number of frames (before subsampling): {c.num_frames}. "
+            #         f"Number of frames (after subsampling): {T}. "
+            #         f"Text: {c.supervisions[0].text}. "
+            #         f"Tokens: {tokens}. "
+            #         f"Number of tokens: {len(tokens)}"
+            #     )
+            return False
+
+        return True
+
+    # train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = aishell.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = aishell.valid_cuts()
+    valid_dl = aishell.valid_dataloaders(valid_cuts)
+
+    # if not params.print_diagnostics:
+    #     scan_pessimistic_batches_for_oom(
+    #         model=model,
+    #         train_dl=train_dl,
+    #         optimizer=optimizer,
+    #         graph_compiler=graph_compiler,
+    #         params=params,
+    #     )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train2.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train2.py
new file mode 100755
index 000000000..88eb34104
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train2.py
@@ -0,0 +1,1253 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --max-duration 300
+
+# For mix precision training:
+
+./pruned_transducer_stateless7_streaming/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless7_streaming/exp \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import AishellAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer2 import Zipformer
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for module in model.modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,4,3,2,4",
+        help="Number of zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dims",
+        type=str,
+        default="1024,1024,2048,2048,1024",
+        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=str,
+        default="8,8,8,8,8",
+        help="Number of attention heads in the zipformer encoder layers.",
+    )
+
+    parser.add_argument(
+        "--encoder-dims",
+        type=str,
+        default="384,384,384,384,384",
+        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
+    )
+
+    parser.add_argument(
+        "--attention-dims",
+        type=str,
+        default="192,192,192,192,192",
+        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
+        not the same as embedding dimension.""",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dims",
+        type=str,
+        default="256,256,256,256,256",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
+        " worse.",
+    )
+
+    parser.add_argument(
+        "--zipformer-downsampling-factors",
+        type=str,
+        default="1,2,4,8,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernels",
+        type=str,
+        default="31,31,31,31,31",
+        help="Sizes of kernels in convolution modules",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--short-chunk-size",
+        type=int,
+        default=50,
+        help="""Chunk length of dynamic training, the chunk size would be either
+        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
+        """,
+    )
+
+    parser.add_argument(
+        "--num-left-chunks",
+        type=int,
+        default=4,
+        help="How many left context can be seen in chunks when calculating attention.",
+    )
+
+    parser.add_argument(
+        "--decode-chunk-len",
+        type=int,
+        default=32,
+        help="The chunk size for decoding (in frames before subsampling)",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7_streaming/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.05, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=2000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Zipformer and Transformer
+    def to_int_tuple(s: str):
+        return tuple(map(int, s.split(",")))
+
+    encoder = Zipformer(
+        num_features=params.feature_dim,
+        output_downsampling_factor=2,
+        zipformer_downsampling_factors=to_int_tuple(
+            params.zipformer_downsampling_factors
+        ),
+        encoder_dims=to_int_tuple(params.encoder_dims),
+        attention_dim=to_int_tuple(params.attention_dims),
+        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
+        nhead=to_int_tuple(params.nhead),
+        feedforward_dim=to_int_tuple(params.feedforward_dims),
+        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
+        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
+        num_left_chunks=params.num_left_chunks,
+        short_chunk_size=params.short_chunk_size,
+        decode_chunk_size=params.decode_chunk_len // 2,
+        is_pnnx=True,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(params.encoder_dims.split(",")[-1]),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute transducer loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            set_batch_count(model, params.batch_idx_train)
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    parameters_names = []
+    parameters_names.append(
+        [name_param_pair[0] for name_param_pair in model.named_parameters()]
+    )
+    optimizer = ScaledAdam(
+        model.parameters(),
+        lr=params.base_lr,
+        clipping_scale=2.0,
+        parameters_names=parameters_names,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            512
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    aishell = AishellAsrDataModule(args)
+
+    train_cuts = aishell.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        # T = ((c.num_frames - 7) // 2 + 1) // 2
+        # tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        # if T < len(tokens):
+        #     logging.warning(
+        #         f"Exclude cut with ID {c.id} from training. "
+        #         f"Number of frames (before subsampling): {c.num_frames}. "
+        #         f"Number of frames (after subsampling): {T}. "
+        #         f"Text: {c.supervisions[0].text}. "
+        #         f"Tokens: {tokens}. "
+        #         f"Number of tokens: {len(tokens)}"
+        #     )
+        #     return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = aishell.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = aishell.valid_cuts()
+    valid_dl = aishell.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    AishellAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer.py
new file mode 120000
index 000000000..ec183baa7
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer2.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer2.py
new file mode 120000
index 000000000..12dbda888
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/zipformer2.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless7_streaming/zipformer2.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py b/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py
index 180930747..6abe6c084 100644
--- a/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py
+++ b/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py
@@ -198,7 +198,7 @@ class AishellAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/asr_datamodule.py b/egs/aishell/ASR/transducer_stateless_modified-2/asr_datamodule.py
index 9c6021a19..cd8dd821c 100644
--- a/egs/aishell/ASR/transducer_stateless_modified-2/asr_datamodule.py
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/asr_datamodule.py
@@ -160,7 +160,7 @@ class AsrDataModule:
         if cuts_musan is not None:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py
index af37cc175..8f6a88f59 100644
--- a/egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py
@@ -218,7 +218,7 @@ class AiShell2AsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/train.py b/egs/aishell2/ASR/pruned_transducer_stateless5/train.py
index 74bf68ccb..8c7448d4c 100755
--- a/egs/aishell2/ASR/pruned_transducer_stateless5/train.py
+++ b/egs/aishell2/ASR/pruned_transducer_stateless5/train.py
@@ -730,7 +730,6 @@ def train_one_epoch(
     tot_loss = MetricsTracker()
 
     for batch_idx, batch in enumerate(train_dl):
-
         params.batch_idx_train += 1
         batch_size = len(batch["supervisions"]["text"])
 
@@ -919,7 +918,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/aishell4/ASR/pruned_transducer_stateless5/asr_datamodule.py
index 835086eb9..e6db2651f 100644
--- a/egs/aishell4/ASR/pruned_transducer_stateless5/asr_datamodule.py
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/asr_datamodule.py
@@ -228,7 +228,7 @@ class Aishell4AsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/train.py b/egs/aishell4/ASR/pruned_transducer_stateless5/train.py
index 47015cbe7..a354f761e 100755
--- a/egs/aishell4/ASR/pruned_transducer_stateless5/train.py
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/train.py
@@ -908,7 +908,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py
index 4799da19d..5ad80817a 100644
--- a/egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py
@@ -211,7 +211,7 @@ class AlimeetingAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/train.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/train.py
index e57b5c859..30154291d 100644
--- a/egs/alimeeting/ASR/pruned_transducer_stateless2/train.py
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/train.py
@@ -635,7 +635,6 @@ def train_one_epoch(
     tot_loss = MetricsTracker()
 
     for batch_idx, batch in enumerate(train_dl):
-
         params.batch_idx_train += 1
         batch_size = len(batch["supervisions"]["text"])
 
@@ -800,7 +799,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/asr_datamodule.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/asr_datamodule.py
index 1cfd053c7..9d288218a 100644
--- a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/asr_datamodule.py
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/asr_datamodule.py
@@ -208,7 +208,7 @@ class AlimeetingAsrDataModule:
             logging.info("Enable MUSAN")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
@@ -288,7 +288,6 @@ class AlimeetingAsrDataModule:
         return train_dl
 
     def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
-
         transforms = []
         if self.args.concatenate_cuts:
             transforms = [
diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/train.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/train.py
index 45d777922..8f09f1aa5 100755
--- a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/train.py
+++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/train.py
@@ -999,7 +999,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/asr_datamodule.py b/egs/ami/ASR/pruned_transducer_stateless7/asr_datamodule.py
index f7ee9c962..79474f1d8 100644
--- a/egs/ami/ASR/pruned_transducer_stateless7/asr_datamodule.py
+++ b/egs/ami/ASR/pruned_transducer_stateless7/asr_datamodule.py
@@ -214,7 +214,7 @@ class AmiAsrDataModule:
             logging.info("Enable MUSAN")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/ami/ASR/pruned_transducer_stateless7/train.py b/egs/ami/ASR/pruned_transducer_stateless7/train.py
index 8c8d9593b..9b67141c0 100755
--- a/egs/ami/ASR/pruned_transducer_stateless7/train.py
+++ b/egs/ami/ASR/pruned_transducer_stateless7/train.py
@@ -988,7 +988,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py b/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py
index 3dd786d33..1549c1631 100644
--- a/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py
+++ b/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py
@@ -202,7 +202,7 @@ class AmiAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py
index 73f2f1dce..546e9f9dd 100644
--- a/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py
@@ -230,7 +230,7 @@ class CommonVoiceAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/train.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/train.py
index 4bd5b83a2..4aedeffe4 100755
--- a/egs/commonvoice/ASR/pruned_transducer_stateless7/train.py
+++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/train.py
@@ -1019,7 +1019,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/csj/ASR/local/utils/asr_datamodule.py b/egs/csj/ASR/local/utils/asr_datamodule.py
index 272486227..042b6ecbf 100644
--- a/egs/csj/ASR/local/utils/asr_datamodule.py
+++ b/egs/csj/ASR/local/utils/asr_datamodule.py
@@ -256,7 +256,7 @@ class CSJAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.musan_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/train.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/train.py
index 18cb75c37..73fcd67aa 100755
--- a/egs/csj/ASR/pruned_transducer_stateless7_streaming/train.py
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/train.py
@@ -1074,7 +1074,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/train2.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/train2.py
index bc4bcf253..4c866ddd8 100755
--- a/egs/csj/ASR/pruned_transducer_stateless7_streaming/train2.py
+++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/train2.py
@@ -1075,7 +1075,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/gigaspeech/ASR/README.md b/egs/gigaspeech/ASR/README.md
index 32a0457c6..f0d60898c 100644
--- a/egs/gigaspeech/ASR/README.md
+++ b/egs/gigaspeech/ASR/README.md
@@ -15,6 +15,7 @@ ln -sfv /path/to/GigaSpeech download/GigaSpeech
 ## Performance Record
 |                                |  Dev  | Test  |
 |--------------------------------|-------|-------|
+|           `zipformer`          | 10.25 | 10.38 |
 |         `conformer_ctc`        | 10.47 | 10.58 |
 | `pruned_transducer_stateless2` | 10.40 | 10.51 |
 
diff --git a/egs/gigaspeech/ASR/RESULTS.md b/egs/gigaspeech/ASR/RESULTS.md
index 7ab565844..841ebdcfa 100644
--- a/egs/gigaspeech/ASR/RESULTS.md
+++ b/egs/gigaspeech/ASR/RESULTS.md
@@ -1,4 +1,78 @@
 ## Results
+### zipformer (zipformer + pruned stateless transducer)
+
+See <https://github.com/k2-fsa/icefall/pull/1254> for more details.
+
+[zipformer](./zipformer)
+
+- Non-streaming
+- normal-scaled model, number of model parameters: 65549011, i.e., 65.55 M
+
+You can find a pretrained model, training logs, decoding logs, and decoding results at:
+<https://huggingface.co/yfyeung/icefall-asr-gigaspeech-zipformer-2023-10-17>
+
+The tensorboard log for training is available at
+<https://wandb.ai/yifanyeung/icefall-asr-gigaspeech-zipformer-2023-10-20>
+
+You can use <https://github.com/k2-fsa/sherpa> to deploy it.
+
+| decoding method      | test-clean | test-other | comment            |
+|----------------------|------------|------------|--------------------|
+| greedy_search        | 10.31      | 10.50      | --epoch 30 --avg 9 |
+| modified_beam_search | 10.25      | 10.38      | --epoch 30 --avg 9 |
+| fast_beam_search     | 10.26      | 10.48      | --epoch 30 --avg 9 |
+
+The training command is:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --causal 0 \
+  --subset XL \
+  --max-duration 700 \
+  --use-transducer 1 \
+  --use-ctc 0 \
+  --lr-epochs 1 \
+  --master-port 12345
+```
+
+The decoding command is:
+```bash
+export CUDA_VISIBLE_DEVICES=0
+
+# greedy search
+./zipformer/decode.py \
+  --epoch 30 \
+  --avg 9 \
+  --exp-dir ./zipformer/exp \
+  --max-duration 1000 \
+  --decoding-method greedy_search
+
+# modified beam search
+./zipformer/decode.py \
+  --epoch 30 \
+  --avg 9 \
+  --exp-dir ./zipformer/exp \
+  --max-duration 1000 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+
+# fast beam search (one best)
+./zipformer/decode.py \
+  --epoch 30 \
+  --avg 9 \
+  --exp-dir ./zipformer/exp \
+  --max-duration 1000 \
+  --decoding-method fast_beam_search \
+  --beam 20.0 \
+  --max-contexts 8 \
+  --max-states 64
+```
+
 ### GigaSpeech BPE training results (Pruned Transducer 2)
 
 #### 2022-05-12
diff --git a/egs/gigaspeech/ASR/conformer_ctc/asr_datamodule.py b/egs/gigaspeech/ASR/conformer_ctc/asr_datamodule.py
index 9d6e3c42a..a93e224d5 100644
--- a/egs/gigaspeech/ASR/conformer_ctc/asr_datamodule.py
+++ b/egs/gigaspeech/ASR/conformer_ctc/asr_datamodule.py
@@ -194,7 +194,7 @@ class GigaSpeechAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
index 29e72b408..b5b27ce95 100644
--- a/egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
@@ -217,7 +217,7 @@ class GigaSpeechAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/gigaspeech/ASR/zipformer/asr_datamodule.py b/egs/gigaspeech/ASR/zipformer/asr_datamodule.py
new file mode 100644
index 000000000..c4472ed23
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/asr_datamodule.py
@@ -0,0 +1,444 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2023  Xiaomi Corporation     (Author: Yifan Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import glob
+import inspect
+import logging
+import re
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import lhotse
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import AudioSamples, OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class GigaSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+        # GigaSpeech specific arguments
+        group.add_argument(
+            "--subset",
+            type=str,
+            default="XL",
+            help="Select the GigaSpeech subset (XS|S|M|L|XL)",
+        )
+        group.add_argument(
+            "--small-dev",
+            type=str2bool,
+            default=False,
+            help="Should we use only 1000 utterances for dev (speeds up training)",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info(f"About to get train {self.args.subset} cuts")
+        if self.args.subset == "XL":
+            filenames = glob.glob(
+                f"{self.args.manifest_dir}/XL_split/gigaspeech_cuts_XL.*.jsonl.gz"
+            )
+            pattern = re.compile(r"gigaspeech_cuts_XL.([0-9]+).jsonl.gz")
+            idx_filenames = ((int(pattern.search(f).group(1)), f) for f in filenames)
+            idx_filenames = sorted(idx_filenames, key=lambda x: x[0])
+            sorted_filenames = [f[1] for f in idx_filenames]
+            logging.info(
+                f"Loading GigaSpeech {len(sorted_filenames)} splits in lazy mode"
+            )
+            cuts_train = lhotse.combine(
+                lhotse.load_manifest_lazy(p) for p in sorted_filenames
+            )
+        else:
+            path = (
+                self.args.manifest_dir / f"gigaspeech_cuts_{self.args.subset}.jsonl.gz"
+            )
+            cuts_train = CutSet.from_jsonl_lazy(path)
+        return cuts_train
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.manifest_dir / "gigaspeech_cuts_DEV.jsonl.gz"
+        )
+        if self.args.small_dev:
+            return cuts_valid.subset(first=1000)
+        else:
+            return cuts_valid
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "gigaspeech_cuts_TEST.jsonl.gz"
+        )
diff --git a/egs/gigaspeech/ASR/zipformer/beam_search.py b/egs/gigaspeech/ASR/zipformer/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/ctc_decode.py b/egs/gigaspeech/ASR/zipformer/ctc_decode.py
new file mode 100755
index 000000000..aa51036d5
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/ctc_decode.py
@@ -0,0 +1,847 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Liyong Guo,
+#                                                 Quandong Wang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+(1) ctc-decoding
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --decoding-method ctc-decoding
+
+(2) 1best
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --decoding-method 1best
+
+(3) nbest
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --decoding-method nbest
+
+(4) nbest-rescoring
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --nbest-scale 1.0 \
+    --lm-dir data/lm \
+    --decoding-method nbest-rescoring
+
+(5) whole-lattice-rescoring
+./zipformer/ctc_decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --use-ctc 1 \
+    --max-duration 600 \
+    --hlg-scale 0.6 \
+    --nbest-scale 1.0 \
+    --lm-dir data/lm \
+    --decoding-method whole-lattice-rescoring
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import GigaSpeechAsrDataModule
+from train import add_model_arguments, get_params, get_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_n_best_list,
+    rescore_with_whole_lattice,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="ctc-decoding",
+        help="""Decoding method.
+        Supported values are:
+        - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece
+          model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+          It needs neither a lexicon nor an n-gram LM.
+        - (2) 1best. Extract the best path from the decoding lattice as the
+          decoding result.
+        - (3) nbest. Extract n paths from the decoding lattice; the path
+          with the highest score is the decoding result.
+        - (4) nbest-rescoring. Extract n paths from the decoding lattice,
+          rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+          the highest score is the decoding result.
+        - (5) whole-lattice-rescoring. Rescore the decoding lattice with an
+          n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+          is the decoding result.
+          you have trained an RNN LM using ./rnn_lm/train.py
+        - (6) nbest-oracle. Its WER is the lower bound of any n-best
+          rescoring method can achieve. Useful for debugging n-best
+          rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=1.0,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--hlg-scale",
+        type=float,
+        default=0.6,
+        help="""The scale to be applied to `hlg.scores`.
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_decoding_params() -> AttributeDict:
+    """Parameters for decoding."""
+    params = AttributeDict(
+        {
+            "frame_shift_ms": 10,
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+    - key: It indicates the setting used for decoding. For example,
+           if no rescoring is used, the key is the string `no_rescore`.
+           If LM rescoring is used, the key is the string `lm_scale_xxx`,
+           where `xxx` is the value of `lm_scale`. An example key is
+           `lm_scale_0.7`
+    - value: It contains the decoding result. `len(value)` equals to
+             batch size. `value[i]` is the decoding result for the i-th
+             utterance in the given batch.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
+        - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+    ctc_output = model.ctc_output(encoder_out)  # (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                params.subsampling_factor,
+                rounding_mode="floor",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=ctc_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.decoding_method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.decoding_method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<UNK>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.decoding_method in ["1best", "nbest"]:
+        if params.decoding_method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.decoding_method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.decoding_method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.decoding_method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.decoding_method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.decoding_method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.decoding_method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      G:
+        An LM. It is not None when params.decoding_method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(f, f"{test_set_name}-{key}", results)
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    # add decoding params
+    params.update(get_decoding_params())
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "ctc-decoding",
+        "1best",
+        "nbest",
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "nbest-oracle",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    params.vocab_size = num_classes
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = 0
+
+    if params.decoding_method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        HLG.scores *= params.hlg_scale
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.decoding_method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.decoding_method == "whole-lattice-rescoring":
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    test_clean_cuts = gigaspeech.test_clean_cuts()
+    test_other_cuts = gigaspeech.test_other_cuts()
+
+    test_clean_dl = gigaspeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = gigaspeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/zipformer/decode.py b/egs/gigaspeech/ASR/zipformer/decode.py
new file mode 100755
index 000000000..3a0c71484
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/decode.py
@@ -0,0 +1,1065 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search (one best)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./zipformer/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+"""
+
+
+import argparse
+import logging
+import math
+import os
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import GigaSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+    modified_beam_search_lm_rescore,
+    modified_beam_search_lm_rescore_LODR,
+    modified_beam_search_lm_shallow_fusion,
+    modified_beam_search_LODR,
+)
+from gigaspeech_scoring import asr_text_post_processing
+from train import add_model_arguments, get_model, get_params
+
+from icefall import ContextGraph, LmScorer, NgramLm
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - modified_beam_search_LODR
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding-method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding-method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--use-shallow-fusion",
+        type=str2bool,
+        default=False,
+        help="""Use neural network LM for shallow fusion.
+        If you want to use LODR, you will also need to set this to true
+        """,
+    )
+
+    parser.add_argument(
+        "--lm-type",
+        type=str,
+        default="rnn",
+        help="Type of NN lm",
+        choices=["rnn", "transformer"],
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.3,
+        help="""The scale of the neural network LM
+        Used only when `--use-shallow-fusion` is set to True.
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens-ngram",
+        type=int,
+        default=2,
+        help="""The order of the ngram lm.
+        """,
+    )
+
+    parser.add_argument(
+        "--backoff-id",
+        type=int,
+        default=500,
+        help="ID of the backoff symbol in the ngram LM",
+    )
+
+    parser.add_argument(
+        "--context-score",
+        type=float,
+        default=2,
+        help="""
+        The bonus score of each token for the context biasing words/phrases.
+        Used only when --decoding-method is modified_beam_search and
+        modified_beam_search_LODR.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-file",
+        type=str,
+        default="",
+        help="""
+        The path of the context biasing lists, one word/phrase each line
+        Used only when --decoding-method is modified_beam_search and
+        modified_beam_search_LODR.
+        """,
+    )
+    add_model_arguments(parser)
+
+    return parser
+
+
+def post_processing(
+    results: List[Tuple[str, List[str], List[str]]],
+) -> List[Tuple[str, List[str], List[str]]]:
+    new_results = []
+    for key, ref, hyp in results:
+        new_ref = asr_text_post_processing(" ".join(ref)).split()
+        new_hyp = asr_text_post_processing(" ".join(hyp)).split()
+        new_results.append((key, new_ref, new_hyp))
+    return new_results
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+    LM: Optional[LmScorer] = None,
+    ngram_lm=None,
+    ngram_lm_scale: float = 0.0,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding-method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network language model.
+      ngram_lm:
+        A ngram language model
+      ngram_lm_scale:
+        The scale for the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    if params.causal:
+        # this seems to cause insertions at the end of the utterance if used with zipformer.
+        pad_len = 30
+        feature_lens += pad_len
+        feature = torch.nn.functional.pad(
+            feature,
+            pad=(0, 0, 0, pad_len),
+            value=LOG_EPS,
+        )
+
+    encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens)
+
+    hyps = []
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            context_graph=context_graph,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_shallow_fusion":
+        hyp_tokens = modified_beam_search_lm_shallow_fusion(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_LODR":
+        hyp_tokens = modified_beam_search_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LODR_lm=ngram_lm,
+            LODR_lm_scale=ngram_lm_scale,
+            LM=LM,
+            context_graph=context_graph,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search_lm_rescore":
+        lm_scale_list = [0.01 * i for i in range(10, 50)]
+        ans_dict = modified_beam_search_lm_rescore(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.decoding_method == "modified_beam_search_lm_rescore_LODR":
+        lm_scale_list = [0.02 * i for i in range(2, 30)]
+        ans_dict = modified_beam_search_lm_rescore_LODR(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+            LM=LM,
+            LODR_lm=ngram_lm,
+            sp=sp,
+            lm_scale_list=lm_scale_list,
+        )
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
+    elif "modified_beam_search" in params.decoding_method:
+        prefix = f"beam_size_{params.beam_size}"
+        if params.decoding_method in (
+            "modified_beam_search_lm_rescore",
+            "modified_beam_search_lm_rescore_LODR",
+        ):
+            ans = dict()
+            assert ans_dict is not None
+            for key, hyps in ans_dict.items():
+                hyps = [sp.decode(hyp).split() for hyp in hyps]
+                ans[f"{prefix}_{key}"] = hyps
+            return ans
+        else:
+            if params.has_contexts:
+                prefix += f"-context-score-{params.context_score}"
+            return {prefix: hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+    context_graph: Optional[ContextGraph] = None,
+    LM: Optional[LmScorer] = None,
+    ngram_lm=None,
+    ngram_lm_scale: float = 0.0,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding-method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+            word_table=word_table,
+            batch=batch,
+            LM=LM,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = post_processing(results)
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    LmScorer.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
+        "modified_beam_search",
+        "modified_beam_search_LODR",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_lm_rescore",
+        "modified_beam_search_lm_rescore_LODR",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if os.path.exists(params.context_file):
+        params.has_contexts = True
+    else:
+        params.has_contexts = False
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        if params.decoding_method in (
+            "modified_beam_search",
+            "modified_beam_search_LODR",
+        ):
+            if params.has_contexts:
+                params.suffix += f"-context-score-{params.context_score}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_shallow_fusion:
+        params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}"
+
+        if "LODR" in params.decoding_method:
+            params.suffix += (
+                f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}"
+            )
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    # only load the neural network LM if required
+    if params.use_shallow_fusion or params.decoding_method in (
+        "modified_beam_search_lm_rescore",
+        "modified_beam_search_lm_rescore_LODR",
+        "modified_beam_search_lm_shallow_fusion",
+        "modified_beam_search_LODR",
+    ):
+        LM = LmScorer(
+            lm_type=params.lm_type,
+            params=params,
+            device=device,
+            lm_scale=params.lm_scale,
+        )
+        LM.to(device)
+        LM.eval()
+    else:
+        LM = None
+
+    # only load N-gram LM when needed
+    if params.decoding_method == "modified_beam_search_lm_rescore_LODR":
+        try:
+            import kenlm
+        except ImportError:
+            print("Please install kenlm first. You can use")
+            print(" pip install https://github.com/kpu/kenlm/archive/master.zip")
+            print("to install it")
+            import sys
+
+            sys.exit(-1)
+        ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa")
+        logging.info(f"lm filename: {ngram_file_name}")
+        ngram_lm = kenlm.Model(ngram_file_name)
+        ngram_lm_scale = None  # use a list to search
+
+    elif params.decoding_method == "modified_beam_search_LODR":
+        lm_filename = f"{params.tokens_ngram}gram.fst.txt"
+        logging.info(f"Loading token level lm: {lm_filename}")
+        ngram_lm = NgramLm(
+            str(params.lang_dir / lm_filename),
+            backoff_id=params.backoff_id,
+            is_binary=False,
+        )
+        logging.info(f"num states: {ngram_lm.lm.num_states}")
+        ngram_lm_scale = params.ngram_lm_scale
+    else:
+        ngram_lm = None
+        ngram_lm_scale = None
+
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+        word_table = None
+
+    if "modified_beam_search" in params.decoding_method:
+        if os.path.exists(params.context_file):
+            contexts = []
+            for line in open(params.context_file).readlines():
+                contexts.append(line.strip())
+            context_graph = ContextGraph(params.context_score)
+            context_graph.build(sp.encode(contexts))
+        else:
+            context_graph = None
+    else:
+        context_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    dev_cuts = gigaspeech.dev_cuts()
+    test_cuts = gigaspeech.test_cuts()
+
+    dev_dl = gigaspeech.test_dataloaders(dev_cuts)
+    test_dl = gigaspeech.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dls = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+            context_graph=context_graph,
+            LM=LM,
+            ngram_lm=ngram_lm,
+            ngram_lm_scale=ngram_lm_scale,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/zipformer/decode_stream.py b/egs/gigaspeech/ASR/zipformer/decode_stream.py
new file mode 120000
index 000000000..b8d8ddfc4
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/decode_stream.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decode_stream.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/decoder.py b/egs/gigaspeech/ASR/zipformer/decoder.py
new file mode 120000
index 000000000..5a8018680
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decoder.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/encoder_interface.py b/egs/gigaspeech/ASR/zipformer/encoder_interface.py
new file mode 120000
index 000000000..653c5b09a
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/export-onnx-ctc.py b/egs/gigaspeech/ASR/zipformer/export-onnx-ctc.py
new file mode 120000
index 000000000..f9d756352
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/export-onnx-ctc.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx-ctc.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/export-onnx-streaming.py b/egs/gigaspeech/ASR/zipformer/export-onnx-streaming.py
new file mode 120000
index 000000000..2962eb784
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/export-onnx-streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/export-onnx-streaming.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/export-onnx.py b/egs/gigaspeech/ASR/zipformer/export-onnx.py
new file mode 100755
index 000000000..0f78cfe5b
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/export-onnx.py
@@ -0,0 +1,620 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang, Wei Kang)
+# Copyright 2023 Danqing Fu (danqing.fu@gmail.com)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/yfyeung/icefall-asr-gigaspeech-zipformer-2023-10-17
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/gigaspeech/ASR
+
+repo_url=https://huggingface.co/yfyeung/icefall-asr-gigaspeech-zipformer-2023-10-17
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./zipformer/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers "2,2,3,4,3,2" \
+  --downsampling-factor "1,2,4,8,4,2" \
+  --feedforward-dim "512,768,1024,1536,1024,768" \
+  --num-heads "4,4,4,8,4,4" \
+  --encoder-dim "192,256,384,512,384,256" \
+  --query-head-dim 32 \
+  --value-head-dim 12 \
+  --pos-head-dim 4 \
+  --pos-dim 48 \
+  --encoder-unmasked-dim "192,192,256,256,256,192" \
+  --cnn-module-kernel "31,31,15,15,15,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --causal False \
+  --chunk-size "16,32,64,-1" \
+  --left-context-frames "64,128,256,-1"
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_model, get_params
+from zipformer import Zipformer2
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import make_pad_mask, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Zipformer and the encoder_proj from the joiner"""
+
+    def __init__(
+        self, encoder: Zipformer2, encoder_embed: nn.Module, encoder_proj: nn.Linear
+    ):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Zipformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        x, x_lens = self.encoder_embed(x, x_lens)
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask)
+        encoder_out = encoder_out.permute(1, 0, 2)
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    encoder_model = torch.jit.trace(encoder_model, (x, x_lens))
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "zipformer2",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "non-streaming zipformer2",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True, is_onnx=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_embed=model.encoder_embed,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/gigaspeech/ASR/zipformer/export.py b/egs/gigaspeech/ASR/zipformer/export.py
new file mode 100755
index 000000000..e45c96b57
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/export.py
@@ -0,0 +1,522 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+
+Usage:
+
+Note: This is a example for gigaspeech dataset, if you are using different
+dataset, you should change the argument values according to your dataset.
+
+(1) Export to torchscript model using torch.jit.script()
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `jit_script.pt` in the given `exp_dir`. You can later
+load it by `torch.jit.load("jit_script.pt")`.
+
+Check ./jit_pretrained.py for its usage.
+
+Check https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --chunk-size 16 \
+  --left-context-frames 128 \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9 \
+  --jit 1
+
+It will generate a file `jit_script_chunk_16_left_128.pt` in the given `exp_dir`.
+You can later load it by `torch.jit.load("jit_script_chunk_16_left_128.pt")`.
+
+Check ./jit_pretrained_streaming.py for its usage.
+
+Check https://github.com/k2-fsa/sherpa
+for how to use the exported models outside of icefall.
+
+(2) Export `model.state_dict()`
+
+- For non-streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9
+
+- For streaming model:
+
+./zipformer/export.py \
+  --exp-dir ./zipformer/exp \
+  --causal 1 \
+  --tokens data/lang_bpe_500/tokens.txt \
+  --epoch 30 \
+  --avg 9
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+- For non-streaming model:
+
+To use the generated file with `zipformer/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/gigaspeech/ASR
+    ./zipformer/decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+- For streaming model:
+
+To use the generated file with `zipformer/decode.py` and `zipformer/streaming_decode.py`, you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/gigaspeech/ASR
+
+    # simulated streaming decoding
+    ./zipformer/decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --causal 1 \
+        --chunk-size 16 \
+        --left-context-frames 128 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+    # chunk-wise streaming decoding
+    ./zipformer/streaming_decode.py \
+        --exp-dir ./zipformer/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --causal 1 \
+        --chunk-size 16 \
+        --left-context-frames 128 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+
+Check ./pretrained.py for its usage.
+
+Note: If you don't want to train a model from scratch, we have
+provided one for you. You can get it at
+
+- non-streaming model:
+https://huggingface.co/yfyeung/icefall-asr-gigaspeech-zipformer-2023-10-17
+
+with the following commands:
+
+    sudo apt-get install git-lfs
+    git lfs install
+    git clone https://huggingface.co/yfyeung/icefall-asr-gigaspeech-zipformer-2023-10-17
+    # You will find the pre-trained models in exp dir
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+import k2
+import torch
+from scaling_converter import convert_scaled_to_non_scaled
+from torch import Tensor, nn
+from train import add_model_arguments, get_model, get_params
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import make_pad_mask, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named jit_script.pt.
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+class EncoderModel(nn.Module):
+    """A wrapper for encoder and encoder_embed"""
+
+    def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None:
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+
+    def forward(
+        self, features: Tensor, feature_lengths: Tensor
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+            features: (N, T, C)
+            feature_lengths: (N,)
+        """
+        x, x_lens = self.encoder_embed(features, feature_lengths)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask)
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return encoder_out, encoder_out_lens
+
+
+class StreamingEncoderModel(nn.Module):
+    """A wrapper for encoder and encoder_embed"""
+
+    def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None:
+        super().__init__()
+        assert len(encoder.chunk_size) == 1, encoder.chunk_size
+        assert len(encoder.left_context_frames) == 1, encoder.left_context_frames
+        self.chunk_size = encoder.chunk_size[0]
+        self.left_context_len = encoder.left_context_frames[0]
+
+        # The encoder_embed subsample features (T - 7) // 2
+        # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+        self.pad_length = 7 + 2 * 3
+
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+
+    def forward(
+        self, features: Tensor, feature_lengths: Tensor, states: List[Tensor]
+    ) -> Tuple[Tensor, Tensor, List[Tensor]]:
+        """Streaming forward for encoder_embed and encoder.
+
+        Args:
+            features: (N, T, C)
+            feature_lengths: (N,)
+            states: a list of Tensors
+
+        Returns encoder outputs, output lengths, and updated states.
+        """
+        chunk_size = self.chunk_size
+        left_context_len = self.left_context_len
+
+        cached_embed_left_pad = states[-2]
+        x, x_lens, new_cached_embed_left_pad = self.encoder_embed.streaming_forward(
+            x=features,
+            x_lens=feature_lengths,
+            cached_left_pad=cached_embed_left_pad,
+        )
+        assert x.size(1) == chunk_size, (x.size(1), chunk_size)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+
+        # processed_mask is used to mask out initial states
+        processed_mask = torch.arange(left_context_len, device=x.device).expand(
+            x.size(0), left_context_len
+        )
+        processed_lens = states[-1]  # (batch,)
+        # (batch, left_context_size)
+        processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+        # Update processed lengths
+        new_processed_lens = processed_lens + x_lens
+
+        # (batch, left_context_size + chunk_size)
+        src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        encoder_states = states[:-2]
+
+        (
+            encoder_out,
+            encoder_out_lens,
+            new_encoder_states,
+        ) = self.encoder.streaming_forward(
+            x=x,
+            x_lens=x_lens,
+            states=encoder_states,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        new_states = new_encoder_states + [
+            new_cached_embed_left_pad,
+            new_processed_lens,
+        ]
+        return encoder_out, encoder_out_lens, new_states
+
+    @torch.jit.export
+    def get_init_states(
+        self,
+        batch_size: int = 1,
+        device: torch.device = torch.device("cpu"),
+    ) -> List[torch.Tensor]:
+        """
+        Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+        is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+        states[-2] is the cached left padding for ConvNeXt module,
+        of shape (batch_size, num_channels, left_pad, num_freqs)
+        states[-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+        """
+        states = self.encoder.get_init_states(batch_size, device)
+
+        embed_states = self.encoder_embed.get_init_states(batch_size, device)
+        states.append(embed_states)
+
+        processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
+        states.append(processed_lens)
+
+        return states
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    # if torch.cuda.is_available():
+    #     device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    if params.jit is True:
+        convert_scaled_to_non_scaled(model, inplace=True)
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+
+        # Wrap encoder and encoder_embed as a module
+        if params.causal:
+            model.encoder = StreamingEncoderModel(model.encoder, model.encoder_embed)
+            chunk_size = model.encoder.chunk_size
+            left_context_len = model.encoder.left_context_len
+            filename = f"jit_script_chunk_{chunk_size}_left_{left_context_len}.pt"
+        else:
+            model.encoder = EncoderModel(model.encoder, model.encoder_embed)
+            filename = "jit_script.pt"
+
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        model.save(str(params.exp_dir / filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torchscript. Export model.state_dict()")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/gigaspeech/ASR/zipformer/gigaspeech_scoring.py b/egs/gigaspeech/ASR/zipformer/gigaspeech_scoring.py
new file mode 120000
index 000000000..a6a4d12b1
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/gigaspeech_scoring.py
@@ -0,0 +1 @@
+../conformer_ctc/gigaspeech_scoring.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/jit_pretrained.py b/egs/gigaspeech/ASR/zipformer/jit_pretrained.py
new file mode 120000
index 000000000..25108391f
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/jit_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/jit_pretrained_ctc.py b/egs/gigaspeech/ASR/zipformer/jit_pretrained_ctc.py
new file mode 120000
index 000000000..9a8da5844
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/jit_pretrained_ctc.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained_ctc.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/jit_pretrained_streaming.py b/egs/gigaspeech/ASR/zipformer/jit_pretrained_streaming.py
new file mode 120000
index 000000000..1962351e9
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/jit_pretrained_streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/jit_pretrained_streaming.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/joiner.py b/egs/gigaspeech/ASR/zipformer/joiner.py
new file mode 120000
index 000000000..5b8a36332
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/joiner.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/model.py b/egs/gigaspeech/ASR/zipformer/model.py
new file mode 120000
index 000000000..cd7e07d72
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/model.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/model.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_check.py b/egs/gigaspeech/ASR/zipformer/onnx_check.py
new file mode 120000
index 000000000..f3dd42004
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_check.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_check.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_decode.py b/egs/gigaspeech/ASR/zipformer/onnx_decode.py
new file mode 120000
index 000000000..0573b88c5
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_decode.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_decode.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_pretrained-streaming.py b/egs/gigaspeech/ASR/zipformer/onnx_pretrained-streaming.py
new file mode 120000
index 000000000..cfea104c2
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_pretrained-streaming.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained-streaming.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_pretrained.py b/egs/gigaspeech/ASR/zipformer/onnx_pretrained.py
new file mode 120000
index 000000000..8f32f4ee7
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc.py b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc.py
new file mode 120000
index 000000000..a3183ebf6
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained_ctc.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_H.py b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_H.py
new file mode 120000
index 000000000..a4fd76ac2
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_H.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained_ctc_H.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_HL.py b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_HL.py
new file mode 120000
index 000000000..f805e3761
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_HL.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained_ctc_HL.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py
new file mode 120000
index 000000000..8343d5079
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/optim.py b/egs/gigaspeech/ASR/zipformer/optim.py
new file mode 120000
index 000000000..5eaa3cffd
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/optim.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/pretrained.py b/egs/gigaspeech/ASR/zipformer/pretrained.py
new file mode 120000
index 000000000..0bd71dde4
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/pretrained.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/pretrained_ctc.py b/egs/gigaspeech/ASR/zipformer/pretrained_ctc.py
new file mode 120000
index 000000000..c2f6f6fc3
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/pretrained_ctc.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/pretrained_ctc.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/profile.py b/egs/gigaspeech/ASR/zipformer/profile.py
new file mode 120000
index 000000000..c93adbd14
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/profile.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/profile.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/scaling.py b/egs/gigaspeech/ASR/zipformer/scaling.py
new file mode 120000
index 000000000..6f398f431
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/scaling_converter.py b/egs/gigaspeech/ASR/zipformer/scaling_converter.py
new file mode 120000
index 000000000..b0ecee05e
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/scaling_converter.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling_converter.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/streaming_beam_search.py b/egs/gigaspeech/ASR/zipformer/streaming_beam_search.py
new file mode 120000
index 000000000..b1ed54557
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/streaming_beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/streaming_beam_search.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/streaming_decode.py b/egs/gigaspeech/ASR/zipformer/streaming_decode.py
new file mode 100755
index 000000000..a76788859
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/streaming_decode.py
@@ -0,0 +1,853 @@
+#!/usr/bin/env python3
+# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang,
+#                                                  Fangjun Kuang,
+#                                                  Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+./zipformer/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --causal 1 \
+  --chunk-size 32 \
+  --left-context-frames 256 \
+  --exp-dir ./zipformer/exp \
+  --decoding-method greedy_search \
+  --num-decode-streams 2000
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import numpy as np
+import sentencepiece as spm
+import torch
+from asr_datamodule import GigaSpeechAsrDataModule
+from decode_stream import DecodeStream
+from kaldifeat import Fbank, FbankOptions
+from lhotse import CutSet
+from streaming_beam_search import (
+    fast_beam_search_one_best,
+    greedy_search,
+    modified_beam_search,
+)
+from torch import Tensor, nn
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    make_pad_mask,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Supported decoding methods are:
+        greedy_search
+        modified_beam_search
+        fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--num_active_paths",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=32,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--num-decode-streams",
+        type=int,
+        default=2000,
+        help="The number of streams that can be decoded parallel.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_init_states(
+    model: nn.Module,
+    batch_size: int = 1,
+    device: torch.device = torch.device("cpu"),
+) -> List[torch.Tensor]:
+    """
+    Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+    is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+    states[-2] is the cached left padding for ConvNeXt module,
+    of shape (batch_size, num_channels, left_pad, num_freqs)
+    states[-1] is processed_lens of shape (batch,), which records the number
+    of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+    """
+    states = model.encoder.get_init_states(batch_size, device)
+
+    embed_states = model.encoder_embed.get_init_states(batch_size, device)
+    states.append(embed_states)
+
+    processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device)
+    states.append(processed_lens)
+
+    return states
+
+
+def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]:
+    """Stack list of zipformer states that correspond to separate utterances
+    into a single emformer state, so that it can be used as an input for
+    zipformer when those utterances are formed into a batch.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance. For element-n,
+        state_list[n] is a list of cached tensors of all encoder layers. For layer-i,
+        state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1,
+        cached_val2, cached_conv1, cached_conv2).
+        state_list[n][-2] is the cached left padding for ConvNeXt module,
+          of shape (batch_size, num_channels, left_pad, num_freqs)
+        state_list[n][-1] is processed_lens of shape (batch,), which records the number
+        of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+    """
+    batch_size = len(state_list)
+    assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0])
+    tot_num_layers = (len(state_list[0]) - 2) // 6
+
+    batch_states = []
+    for layer in range(tot_num_layers):
+        layer_offset = layer * 6
+        # cached_key: (left_context_len, batch_size, key_dim)
+        cached_key = torch.cat(
+            [state_list[i][layer_offset] for i in range(batch_size)], dim=1
+        )
+        # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
+        cached_nonlin_attn = torch.cat(
+            [state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1
+        )
+        # cached_val1: (left_context_len, batch_size, value_dim)
+        cached_val1 = torch.cat(
+            [state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1
+        )
+        # cached_val2: (left_context_len, batch_size, value_dim)
+        cached_val2 = torch.cat(
+            [state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1
+        )
+        # cached_conv1: (#batch, channels, left_pad)
+        cached_conv1 = torch.cat(
+            [state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0
+        )
+        # cached_conv2: (#batch, channels, left_pad)
+        cached_conv2 = torch.cat(
+            [state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0
+        )
+        batch_states += [
+            cached_key,
+            cached_nonlin_attn,
+            cached_val1,
+            cached_val2,
+            cached_conv1,
+            cached_conv2,
+        ]
+
+    cached_embed_left_pad = torch.cat(
+        [state_list[i][-2] for i in range(batch_size)], dim=0
+    )
+    batch_states.append(cached_embed_left_pad)
+
+    processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0)
+    batch_states.append(processed_lens)
+
+    return batch_states
+
+
+def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]:
+    """Unstack the zipformer state corresponding to a batch of utterances
+    into a list of states, where the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Note:
+      It is the inverse of :func:`stack_states`.
+
+    Args:
+        batch_states: A list of cached tensors of all encoder layers. For layer-i,
+          states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2,
+          cached_conv1, cached_conv2).
+          state_list[-2] is the cached left padding for ConvNeXt module,
+          of shape (batch_size, num_channels, left_pad, num_freqs)
+          states[-1] is processed_lens of shape (batch,), which records the number
+          of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch.
+
+    Returns:
+        state_list: A list of list. Each element in state_list corresponding to the internal state
+        of the zipformer model for a single utterance.
+    """
+    assert (len(batch_states) - 2) % 6 == 0, len(batch_states)
+    tot_num_layers = (len(batch_states) - 2) // 6
+
+    processed_lens = batch_states[-1]
+    batch_size = processed_lens.shape[0]
+
+    state_list = [[] for _ in range(batch_size)]
+
+    for layer in range(tot_num_layers):
+        layer_offset = layer * 6
+        # cached_key: (left_context_len, batch_size, key_dim)
+        cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1)
+        # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
+        cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_val1: (left_context_len, batch_size, value_dim)
+        cached_val1_list = batch_states[layer_offset + 2].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_val2: (left_context_len, batch_size, value_dim)
+        cached_val2_list = batch_states[layer_offset + 3].chunk(
+            chunks=batch_size, dim=1
+        )
+        # cached_conv1: (#batch, channels, left_pad)
+        cached_conv1_list = batch_states[layer_offset + 4].chunk(
+            chunks=batch_size, dim=0
+        )
+        # cached_conv2: (#batch, channels, left_pad)
+        cached_conv2_list = batch_states[layer_offset + 5].chunk(
+            chunks=batch_size, dim=0
+        )
+        for i in range(batch_size):
+            state_list[i] += [
+                cached_key_list[i],
+                cached_nonlin_attn_list[i],
+                cached_val1_list[i],
+                cached_val2_list[i],
+                cached_conv1_list[i],
+                cached_conv2_list[i],
+            ]
+
+    cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0)
+    for i in range(batch_size):
+        state_list[i].append(cached_embed_left_pad_list[i])
+
+    processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0)
+    for i in range(batch_size):
+        state_list[i].append(processed_lens_list[i])
+
+    return state_list
+
+
+def streaming_forward(
+    features: Tensor,
+    feature_lens: Tensor,
+    model: nn.Module,
+    states: List[Tensor],
+    chunk_size: int,
+    left_context_len: int,
+) -> Tuple[Tensor, Tensor, List[Tensor]]:
+    """
+    Returns encoder outputs, output lengths, and updated states.
+    """
+    cached_embed_left_pad = states[-2]
+    (x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward(
+        x=features,
+        x_lens=feature_lens,
+        cached_left_pad=cached_embed_left_pad,
+    )
+    assert x.size(1) == chunk_size, (x.size(1), chunk_size)
+
+    src_key_padding_mask = make_pad_mask(x_lens)
+
+    # processed_mask is used to mask out initial states
+    processed_mask = torch.arange(left_context_len, device=x.device).expand(
+        x.size(0), left_context_len
+    )
+    processed_lens = states[-1]  # (batch,)
+    # (batch, left_context_size)
+    processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1)
+    # Update processed lengths
+    new_processed_lens = processed_lens + x_lens
+
+    # (batch, left_context_size + chunk_size)
+    src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
+
+    x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+    encoder_states = states[:-2]
+    (
+        encoder_out,
+        encoder_out_lens,
+        new_encoder_states,
+    ) = model.encoder.streaming_forward(
+        x=x,
+        x_lens=x_lens,
+        states=encoder_states,
+        src_key_padding_mask=src_key_padding_mask,
+    )
+    encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+    new_states = new_encoder_states + [
+        new_cached_embed_left_pad,
+        new_processed_lens,
+    ]
+    return encoder_out, encoder_out_lens, new_states
+
+
+def decode_one_chunk(
+    params: AttributeDict,
+    model: nn.Module,
+    decode_streams: List[DecodeStream],
+) -> List[int]:
+    """Decode one chunk frames of features for each decode_streams and
+    return the indexes of finished streams in a List.
+
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      decode_streams:
+        A List of DecodeStream, each belonging to a utterance.
+    Returns:
+      Return a List containing which DecodeStreams are finished.
+    """
+    device = model.device
+    chunk_size = int(params.chunk_size)
+    left_context_len = int(params.left_context_frames)
+
+    features = []
+    feature_lens = []
+    states = []
+    processed_lens = []  # Used in fast-beam-search
+
+    for stream in decode_streams:
+        feat, feat_len = stream.get_feature_frames(chunk_size * 2)
+        features.append(feat)
+        feature_lens.append(feat_len)
+        states.append(stream.states)
+        processed_lens.append(stream.done_frames)
+
+    feature_lens = torch.tensor(feature_lens, device=device)
+    features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS)
+
+    # Make sure the length after encoder_embed is at least 1.
+    # The encoder_embed subsample features (T - 7) // 2
+    # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
+    tail_length = chunk_size * 2 + 7 + 2 * 3
+    if features.size(1) < tail_length:
+        pad_length = tail_length - features.size(1)
+        feature_lens += pad_length
+        features = torch.nn.functional.pad(
+            features,
+            (0, 0, 0, pad_length),
+            mode="constant",
+            value=LOG_EPS,
+        )
+
+    states = stack_states(states)
+
+    encoder_out, encoder_out_lens, new_states = streaming_forward(
+        features=features,
+        feature_lens=feature_lens,
+        model=model,
+        states=states,
+        chunk_size=chunk_size,
+        left_context_len=left_context_len,
+    )
+
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+
+    if params.decoding_method == "greedy_search":
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
+    elif params.decoding_method == "fast_beam_search":
+        processed_lens = torch.tensor(processed_lens, device=device)
+        processed_lens = processed_lens + encoder_out_lens
+        fast_beam_search_one_best(
+            model=model,
+            encoder_out=encoder_out,
+            processed_lens=processed_lens,
+            streams=decode_streams,
+            beam=params.beam,
+            max_states=params.max_states,
+            max_contexts=params.max_contexts,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        modified_beam_search(
+            model=model,
+            streams=decode_streams,
+            encoder_out=encoder_out,
+            num_active_paths=params.num_active_paths,
+        )
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+
+    states = unstack_states(new_states)
+
+    finished_streams = []
+    for i in range(len(decode_streams)):
+        decode_streams[i].states = states[i]
+        decode_streams[i].done_frames += encoder_out_lens[i]
+        if decode_streams[i].done:
+            finished_streams.append(i)
+
+    return finished_streams
+
+
+def decode_dataset(
+    cuts: CutSet,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      cuts:
+        Lhotse Cutset containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    device = model.device
+
+    opts = FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+
+    log_interval = 100
+
+    decode_results = []
+    # Contain decode streams currently running.
+    decode_streams = []
+    for num, cut in enumerate(cuts):
+        # each utterance has a DecodeStream.
+        initial_states = get_init_states(model=model, batch_size=1, device=device)
+        decode_stream = DecodeStream(
+            params=params,
+            cut_id=cut.id,
+            initial_states=initial_states,
+            decoding_graph=decoding_graph,
+            device=device,
+        )
+
+        audio: np.ndarray = cut.load_audio()
+        # audio.shape: (1, num_samples)
+        assert len(audio.shape) == 2
+        assert audio.shape[0] == 1, "Should be single channel"
+        assert audio.dtype == np.float32, audio.dtype
+
+        # The trained model is using normalized samples
+        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+
+        samples = torch.from_numpy(audio).squeeze(0)
+
+        fbank = Fbank(opts)
+        feature = fbank(samples.to(device))
+        decode_stream.set_features(feature, tail_pad_len=30)
+        decode_stream.ground_truth = cut.supervisions[0].text
+
+        decode_streams.append(decode_stream)
+
+        while len(decode_streams) >= params.num_decode_streams:
+            finished_streams = decode_one_chunk(
+                params=params, model=model, decode_streams=decode_streams
+            )
+            for i in sorted(finished_streams, reverse=True):
+                decode_results.append(
+                    (
+                        decode_streams[i].id,
+                        decode_streams[i].ground_truth.split(),
+                        sp.decode(decode_streams[i].decoding_result()).split(),
+                    )
+                )
+                del decode_streams[i]
+
+        if num % log_interval == 0:
+            logging.info(f"Cuts processed until now is {num}.")
+
+    # decode final chunks of last sequences
+    while len(decode_streams):
+        finished_streams = decode_one_chunk(
+            params=params, model=model, decode_streams=decode_streams
+        )
+        for i in sorted(finished_streams, reverse=True):
+            decode_results.append(
+                (
+                    decode_streams[i].id,
+                    decode_streams[i].ground_truth.split(),
+                    sp.decode(decode_streams[i].decoding_result()).split(),
+                )
+            )
+            del decode_streams[i]
+
+    if params.decoding_method == "greedy_search":
+        key = "greedy_search"
+    elif params.decoding_method == "fast_beam_search":
+        key = (
+            f"beam_{params.beam}_"
+            f"max_contexts_{params.max_contexts}_"
+            f"max_states_{params.max_states}"
+        )
+    elif params.decoding_method == "modified_beam_search":
+        key = f"num_active_paths_{params.num_active_paths}"
+    else:
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+    return {key: decode_results}
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    params.res_dir = params.exp_dir / "streaming" / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    assert params.causal, params.causal
+    assert "," not in params.chunk_size, "chunk_size should be one value in decoding."
+    assert (
+        "," not in params.left_context_frames
+    ), "left_context_frames should be one value in decoding."
+    params.suffix += f"-chunk-{params.chunk_size}"
+    params.suffix += f"-left-context-{params.left_context_frames}"
+
+    # for fast_beam_search
+    if params.decoding_method == "fast_beam_search":
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if start >= 0:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    decoding_graph = None
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    dev_cuts = gigaspeech.dev_cuts()
+    test_cuts = gigaspeech.test_cuts()
+
+    test_sets = ["dev", "test"]
+    test_cuts = [dev_cuts, test_cuts]
+
+    for test_set, test_cut in zip(test_sets, test_cuts):
+        results_dict = decode_dataset(
+            cuts=test_cut,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/zipformer/subsampling.py b/egs/gigaspeech/ASR/zipformer/subsampling.py
new file mode 120000
index 000000000..01ae9002c
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/subsampling.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/test_scaling.py b/egs/gigaspeech/ASR/zipformer/test_scaling.py
new file mode 120000
index 000000000..715798436
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/test_scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/test_scaling.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/test_subsampling.py b/egs/gigaspeech/ASR/zipformer/test_subsampling.py
new file mode 120000
index 000000000..bf0ee3d11
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/test_subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/test_subsampling.py
\ No newline at end of file
diff --git a/egs/gigaspeech/ASR/zipformer/train.py b/egs/gigaspeech/ASR/zipformer/train.py
new file mode 100755
index 000000000..d8ff4fecc
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/train.py
@@ -0,0 +1,1345 @@
+#!/usr/bin/env python3
+# Copyright    2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,
+#                                                       Zengwei Yao,
+#                                                       Yifan Yang,
+#                                                       Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+# For non-streaming model training:
+./zipformer/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --max-duration 1000
+
+# For streaming model training:
+./zipformer/train.py \
+  --world-size 8 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp \
+  --causal 1 \
+  --max-duration 1000
+
+It supports training with:
+  - transducer loss (default), with `--use-transducer True --use-ctc False`
+  - ctc loss (not recommended), with `--use-transducer False --use-ctc True`
+  - transducer loss & ctc loss, with `--use-transducer True --use-ctc True`
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import GigaSpeechAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import AsrModel
+from optim import Eden, ScaledAdam
+from scaling import ScheduledFloat
+from subsampling import Conv2dSubsampling
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="Embedding dimension in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="Value dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "A single int or comma-separated list.  Must be <= each corresponding encoder_dim.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="Sizes of convolutional kernels in convolution modules in each encoder stack: "
+        "a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. "
+        " Must be just -1 if --causal=False",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="Maximum left-contexts for causal training, measured in frames which will "
+        "be converted to a number of chunks.  If splitting into chunks, "
+        "chunk left-context frames will be chosen randomly from this list; else not relevant.",
+    )
+
+    parser.add_argument(
+        "--use-transducer",
+        type=str2bool,
+        default=True,
+        help="If True, use Transducer head.",
+    )
+
+    parser.add_argument(
+        "--use-ctc",
+        type=str2bool,
+        default=False,
+        help="If True, use CTC head.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.045, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=1,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="Reference batch duration for purposes of adjusting batch counts for setting various "
+        "schedules inside the model",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--ctc-loss-scale",
+        type=float,
+        default=0.2,
+        help="Scale for CTC loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 1.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 500,
+            "reset_interval": 2000,
+            "valid_interval": 20000,
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_model(params: AttributeDict) -> nn.Module:
+    assert params.use_transducer or params.use_ctc, (
+        f"At least one of them should be True, "
+        f"but got params.use_transducer={params.use_transducer}, "
+        f"params.use_ctc={params.use_ctc}"
+    )
+
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+
+    if params.use_transducer:
+        decoder = get_decoder_model(params)
+        joiner = get_joiner_model(params)
+    else:
+        decoder = None
+        joiner = None
+
+    model = AsrModel(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        vocab_size=params.vocab_size,
+        use_transducer=params.use_transducer,
+        use_ctc=params.use_ctc,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss, ctc_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        loss = 0.0
+
+        if params.use_transducer:
+            s = params.simple_loss_scale
+            # take down the scale on the simple loss from 1.0 at the start
+            # to params.simple_loss scale by warm_step.
+            simple_loss_scale = (
+                s
+                if batch_idx_train >= warm_step
+                else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+            )
+            pruned_loss_scale = (
+                1.0
+                if batch_idx_train >= warm_step
+                else 0.1 + 0.9 * (batch_idx_train / warm_step)
+            )
+            loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+        if params.use_ctc:
+            loss += params.ctc_loss_scale * ctc_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    if params.use_transducer:
+        info["simple_loss"] = simple_loss.detach().cpu().item()
+        info["pruned_loss"] = pruned_loss.detach().cpu().item()
+    if params.use_ctc:
+        info["ctc_loss"] = ctc_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    if not params.use_transducer:
+        params.ctc_loss_scale = 1.0
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            2**22
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    gigaspeech = GigaSpeechAsrDataModule(args)
+
+    train_cuts = gigaspeech.train_cuts()
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = gigaspeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = gigaspeech.dev_cuts()
+    valid_dl = gigaspeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    GigaSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/gigaspeech/ASR/zipformer/zipformer.py b/egs/gigaspeech/ASR/zipformer/zipformer.py
new file mode 120000
index 000000000..23011dda7
--- /dev/null
+++ b/egs/gigaspeech/ASR/zipformer/zipformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/zipformer.py
\ No newline at end of file
diff --git a/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py b/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py
index a72df89e0..c1abdbdb5 100644
--- a/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py
+++ b/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py
@@ -204,7 +204,7 @@ class LibriCssAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/libricss/SURT/prepare.sh b/egs/libricss/SURT/prepare.sh
index 3d2581d96..b2d37f949 100755
--- a/egs/libricss/SURT/prepare.sh
+++ b/egs/libricss/SURT/prepare.sh
@@ -90,6 +90,14 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
   # NOTE: Alignments are required for this recipe.
   mkdir -p data/manifests
 
+  log "This recipe uses mfa alignment for trimming"
+  if [ ! -d $dl_dir/libri_alignments/LibriSpeech ]; then
+    log "No alignment provided. please refer to ../../librispeech/ASR/add_alignments.sh \n \
+        for mfa alignments. Once you have downloaded and unzipped the .zip file containing \n \
+        all alignments, the folder should be renamed to libri_alignments and moved to your $dl_dir ."
+    exit 0
+  fi
+
   lhotse prepare librispeech -p train-clean-100 -p train-clean-360 -p train-other-500 -p dev-clean \
     -j 4 --alignments-dir $dl_dir/libri_alignments/LibriSpeech $dl_dir/librispeech data/manifests/
 fi
@@ -118,9 +126,12 @@ fi
 
 if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
   log "Stage 4: Extract features for LibriSpeech, trim to alignments, and shuffle the cuts"
-  python local/compute_fbank_librispeech.py
-  lhotse combine data/manifests/librispeech_cuts_train* - |\
-    lhotse cut trim-to-alignments --type word --max-pause 0.2 - - |\
+  # python local/compute_fbank_librispeech.py
+  lhotse combine data/manifests/librispeech_cuts_train* data/manifests/librispeech_cuts_train_all.jsonl.gz
+  lhotse cut trim-to-alignments --type word --max-pause 0.2 \
+    data/manifests/librispeech_cuts_train_all.jsonl.gz \
+    data/manifests/librispeech_cuts_train_all_trimmed.jsonl.gz
+  cat <(gunzip -c data/manifests/librispeech_cuts_train_all_trimmed.jsonl.gz) | \
     shuf | gzip -c > data/manifests/librispeech_cuts_train_trimmed.jsonl.gz
 fi
 
@@ -152,7 +163,7 @@ if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
     data/manifests/lsmix_cuts_train_clean_ov40.jsonl.gz
 
   # Full training set (2,3 speakers) anechoic
-  log "Generating anechoic ${part} set (full)"
+  log "Generating anechoic set (full)"
   lhotse workflows simulate-meetings \
     --method conversational \
     --fit-to-supervisions data/manifests/libricss-sdm_supervisions_all_v1.jsonl.gz \
diff --git a/egs/libriheavy/ASR/RESULTS.md b/egs/libriheavy/ASR/RESULTS.md
new file mode 100644
index 000000000..4fbedad98
--- /dev/null
+++ b/egs/libriheavy/ASR/RESULTS.md
@@ -0,0 +1,205 @@
+## Results
+
+### Zipformer PromptASR (zipformer + PromptASR + BERT text encoder)
+
+#### [zipformer_prompt_asr](./zipformer_prompt_asr)
+
+See <https://github.com/k2-fsa/icefall/pull/1250> for commit history and
+our paper <https://arxiv.org/abs/2309.07414> for more details.
+
+
+
+##### Training on the medium subset, with content & style prompt, **no** context list
+
+You can find a pre-trained model, training logs, decoding logs, and decoding results at: <https://huggingface.co/marcoyang/icefall-promptasr-libriheavy-zipformer-BERT-2023-10-10>
+
+The training command is:
+
+```bash
+causal=0
+subset=medium
+memory_dropout_rate=0.05
+text_encoder_type=BERT
+
+python ./zipformer_prompt_asr/train_bert_encoder.py \
+    --world-size 4 \
+    --start-epoch 1 \
+    --num-epochs 60 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --use-fp16 True \
+    --memory-dropout-rate $memory_dropout_rate \
+    --causal $causal \
+    --subset $subset \
+    --manifest-dir data/fbank \
+    --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \
+    --max-duration 1000 \
+    --text-encoder-type $text_encoder_type \
+    --text-encoder-dim 768 \
+    --use-context-list 0 \
+    --top-k $top_k \
+    --use-style-prompt 1
+```
+
+The decoding results using utterance-level context (epoch-60-avg-10):
+
+| decoding method      | lh-test-clean | lh-test-other | comment             |
+|----------------------|---------------|---------------|---------------------|
+| modified_beam_search        |   3.13        |    6.78       |    --use-pre-text False --use-style-prompt False      |
+| modified_beam_search        |   2.86        |    5.93       |    --pre-text-transform upper-no-punc --style-text-transform upper-no-punc       |
+| modified_beam_search        |   2.6        |    5.5       |    --pre-text-transform mixed-punc --style-text-transform mixed-punc       |
+
+
+The decoding command is:
+
+```bash
+for style in mixed-punc upper-no-punc; do
+    python ./zipformer_prompt_asr/decode_bert.py \
+        --epoch 60 \
+        --avg 10 \
+        --use-averaged-model True \
+        --post-normalization True \
+        --causal False \
+        --exp-dir ./zipformer_prompt_asr/exp \
+        --manifest-dir data/fbank \
+        --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \
+        --max-duration 1000 \
+        --decoding-method modified_beam_search \
+        --beam-size 4 \
+        --text-encoder-type BERT \
+        --text-encoder-dim 768 \
+        --memory-layer 0 \
+        --use-ls-test-set False \
+        --use-ls-context-list False \
+        --max-prompt-lens 1000 \
+        --use-pre-text True \
+        --use-style-prompt True \
+        --style-text-transform $style \
+        --pre-text-transform $style \
+        --compute-CER 0
+done
+```
+
+##### Training on the medium subset, with content & style prompt, **with** context list
+
+You can find a pre-trained model, training logs, decoding logs, and decoding results at: <https://huggingface.co/marcoyang/icefall-promptasr-with-context-libriheavy-zipformer-BERT-2023-10-10>
+
+This model is trained with an extra type of content prompt (context words), thus it does better
+on **word-level** context biasing. Note that to train this model, please first run `prepare_prompt_asr.sh`
+to prepare a manifest containing context words.
+
+The training command is:
+
+```bash
+
+causal=0
+subset=medium
+memory_dropout_rate=0.05
+text_encoder_type=BERT
+use_context_list=True
+
+# prepare the required data for context biasing
+./prepare_prompt_asr.sh --stage 0 --stop_stage 1
+
+python ./zipformer_prompt_asr/train_bert_encoder.py \
+    --world-size 4 \
+    --start-epoch 1 \
+    --num-epochs 50 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --use-fp16 True \
+    --memory-dropout-rate $memory_dropout_rate \
+    --causal $causal \
+    --subset $subset \
+    --manifest-dir data/fbank \
+    --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \
+    --max-duration 1000 \
+    --text-encoder-type $text_encoder_type \
+    --text-encoder-dim 768 \
+    --use-context-list $use_context_list \
+    --top-k 10000 \
+    --use-style-prompt 1
+```
+
+*Utterance-level biasing:*
+
+| decoding method      | lh-test-clean | lh-test-other | comment             |
+|----------------------|---------------|---------------|---------------------|
+| modified_beam_search        |   3.17        |    6.72       |    --use-pre-text 0 --use-style-prompt 0    |
+| modified_beam_search        |   2.91        |    6.24       |    --pre-text-transform upper-no-punc --style-text-transform upper-no-punc       |
+| modified_beam_search        |   2.72        |    5.72       |    --pre-text-transform mixed-punc --style-text-transform mixed-punc       |
+
+
+The decoding command for the table above is:
+
+```bash
+for style in mixed-punc upper-no-punc; do
+    python ./zipformer_prompt_asr/decode_bert.py \
+        --epoch 50 \
+        --avg 10 \
+        --use-averaged-model True \
+        --post-normalization True \
+        --causal False \
+        --exp-dir ./zipformer_prompt_asr/exp \
+        --manifest-dir data/fbank \
+        --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \
+        --max-duration 1000 \
+        --decoding-method modified_beam_search \
+        --beam-size 4 \
+        --text-encoder-type BERT \
+        --text-encoder-dim 768 \
+        --memory-layer 0 \
+        --use-ls-test-set False \
+        --use-ls-context-list False \
+        --max-prompt-lens 1000 \
+        --use-pre-text True \
+        --use-style-prompt True \
+        --style-text-transform $style \
+        --pre-text-transform $style \
+        --compute-CER 0
+done
+```
+
+*Word-level biasing:*
+
+The results are reported on LibriSpeech test-sets using the biasing list provided from <https://arxiv.org/abs/2104.02194>.
+You need to set `--use-ls-test-set True` so that the LibriSpeech test sets are used.
+
+| decoding method      | ls-test-clean | ls-test-other | comment             |
+|----------------------|---------------|---------------|---------------------|
+| modified_beam_search        |   2.4        |    5.08       |    --use-pre-text 0 --use-style-prompt 0    |
+| modified_beam_search        |   2.14        |    4.62       |    --use-ls-context-list 1 --pre-text-transform mixed-punc --style-text-transform mixed-punc --ls-distractors 0       |
+| modified_beam_search        |   2.14        |    4.64       |    --use-ls-context-list 1 --pre-text-transform mixed-punc --style-text-transform mixed-punc --ls-distractors 100       |
+
+The decoding command is for the table above is:
+
+```bash
+use_ls_test_set=1
+use_ls_context_list=1
+
+for ls_distractors in 0 100; do
+    python ./zipformer_prompt_asr/decode_bert.py \
+        --epoch 50 \
+        --avg 10 \
+        --use-averaged-model True \
+        --post-normalization True \
+        --causal False \
+        --exp-dir ./zipformer_prompt_asr/exp \
+        --manifest-dir data/fbank \
+        --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \
+        --max-duration 1000 \
+        --decoding-method modified_beam_search \
+        --beam-size 4 \
+        --text-encoder-type BERT \
+        --text-encoder-dim 768 \
+        --memory-layer 0 \
+        --use-ls-test-set $use_ls_test_setse \
+        --use-ls-context-list $use_ls_context_list \
+        --ls-distractors $ls_distractors \
+        --max-prompt-lens 1000 \
+        --use-pre-text True \
+        --use-style-prompt True \
+        --style-text-transform mixed-punc \
+        --pre-text-transform mixed-punc \
+        --compute-CER 0
+done
+
+```
diff --git a/egs/libriheavy/ASR/prepare_prompt_asr.sh b/egs/libriheavy/ASR/prepare_prompt_asr.sh
new file mode 100755
index 000000000..b931cea26
--- /dev/null
+++ b/egs/libriheavy/ASR/prepare_prompt_asr.sh
@@ -0,0 +1,36 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+# This is the preparation recipe for PromptASR: https://arxiv.org/pdf/2309.07414
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+stage=-1
+stop_stage=100
+manifest_dir=data/fbank
+subset=medium
+topk=10000
+
+. shared/parse_options.sh || exit 1
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+    log "Stage 0: Download the meta biasing list for LibriSpeech"
+    mkdir -p data/context_biasing
+    cd data/context_biasing
+    git clone https://github.com/facebookresearch/fbai-speech.git
+    cd ../..
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+    log "Stage 1: Add rare-words for context biasing to the manifest"
+    python zipformer_prompt_asr/utils.py \
+        --manifest-dir $manifest_dir \
+        --subset $subset \
+        --top-k $topk
+
+fi
diff --git a/egs/libriheavy/ASR/shared b/egs/libriheavy/ASR/shared
new file mode 120000
index 000000000..4cbd91a7e
--- /dev/null
+++ b/egs/libriheavy/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/__init__.py b/egs/libriheavy/ASR/zipformer_prompt_asr/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py b/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py
new file mode 100644
index 000000000..690003377
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py
@@ -0,0 +1,520 @@
+# Copyright      2021  Piotr Żelasko
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Callable, Dict, List, Optional
+
+import torch
+from dataset import PromptASRDataset
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # SingleCutSampler,
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    ExtraPadding,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class LibriHeavyAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+        if args.use_context_list:
+            assert args.rare_word_file is not None
+            with open(args.rare_word_file, "r") as f:
+                self.rare_word_list = (
+                    f.read().lower().split()
+                )  # Use lower-cased for easier style transform
+        else:
+            self.rare_word_list = None
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it "
+            "with training dataset. ",
+        )
+
+        # Libriheavy specific arguments
+        group.add_argument(
+            "--subset",
+            type=str,
+            default="small",
+            help="Select the Libriheavy subset (small|medium|large)",
+        )
+
+        group.add_argument(
+            "--use-context-list",
+            type=str2bool,
+            default=False,
+            help="Use the context list of libri heavy",
+        )
+
+        group.add_argument(
+            "--top-k",
+            type=int,
+            default=10000,
+            help="""The top-k words are identified as common words,
+            the rest as rare words""",
+        )
+
+        group.add_argument(
+            "--with-decoding",
+            type=str2bool,
+            default=False,
+            help="If the texts field contain decoding",
+        )
+
+        group.add_argument(
+            "--random-left-padding",
+            type=str2bool,
+        )
+
+        group.add_argument(
+            "--rare-word-file",
+            type=str,
+        )
+
+        group.add_argument(
+            "--long-audio-cuts",
+            type=str,
+            default="data/manifest_npr/npr1_cuts_all_guids_0.jsonl.gz",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+        text_sampling_func: Callable[[List[str]], str] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = PromptASRDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+            text_sampling_func=text_sampling_func,
+            rare_word_list=self.rare_word_list,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = PromptASRDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+                text_sampling_func=text_sampling_func,
+                rare_word_list=self.rare_word_list,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            raise NotImplementedError(
+                "SingleCutSampler is no longer supported by lhotse"
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(
+        self,
+        cuts_valid: CutSet,
+        text_sampling_func: Callable[[List[str]], str] = None,
+    ) -> DataLoader:
+        transforms = []
+        if self.args.random_left_padding:
+            logging.info("Enable random left padding")
+            transforms.append(
+                ExtraPadding(extra_frames=16, randomized=True, direction="left")
+            )
+
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = PromptASRDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+                text_sampling_func=text_sampling_func,
+                rare_word_list=self.rare_word_list,
+            )
+        else:
+            validate = PromptASRDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+                text_sampling_func=text_sampling_func,
+                rare_word_list=self.rare_word_list,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info(f"About to get {self.args.subset} cuts")
+
+        if self.args.use_context_list:
+            path = (
+                self.args.manifest_dir
+                / f"libriheavy_cuts_{self.args.subset}_with_context_list_topk_{self.args.top_k}.jsonl.gz"
+            )
+        elif self.args.with_decoding:
+            path = (
+                self.args.manifest_dir
+                / f"libriheavy_cuts_{self.args.subset}_with_decoding.jsonl.gz"
+            )
+        else:
+            path = (
+                self.args.manifest_dir / f"libriheavy_cuts_{self.args.subset}.jsonl.gz"
+            )
+
+        logging.info(f"Loading manifest from {path}.")
+        cuts_train = CutSet.from_jsonl_lazy(path)
+        return cuts_train
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_dev.jsonl.gz"
+        )
+        return cuts_valid
+
+    @lru_cache()
+    def test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_test-clean_official.jsonl.gz"
+        )
+        return cuts_valid
+
+    @lru_cache()
+    def test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        cuts_valid = load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_test-other_official.jsonl.gz"
+        )
+        return cuts_valid
+
+    @lru_cache()
+    def librispeech_test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def librispeech_test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-other.jsonl.gz"
+        )
+
+    @lru_cache()
+    def long_audio_cuts(self) -> CutSet:
+        logging.info("About to get long audio cuts")
+        cuts = load_manifest_lazy(
+            self.args.long_audio_cuts,
+        )
+        return cuts
+
+    @lru_cache()
+    def test_dev_cuts(self) -> CutSet:
+        logging.info("About to get test dev cuts")
+        cuts = load_manifest_lazy(
+            self.args.manifest_dir / "libriheavy_cuts_test_dev.jsonl.gz"
+        )
+        return cuts
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/beam_search.py b/egs/libriheavy/ASR/zipformer_prompt_asr/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/dataset.py b/egs/libriheavy/ASR/zipformer_prompt_asr/dataset.py
new file mode 100644
index 000000000..e0bf8f73d
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/dataset.py
@@ -0,0 +1,586 @@
+#      Copyright      2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import random
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from lhotse import validate
+from lhotse.cut import CutSet
+from lhotse.dataset import K2SpeechRecognitionDataset
+from lhotse.dataset.input_strategies import BatchIO, PrecomputedFeatures
+from lhotse.utils import compute_num_frames, ifnone
+from text_normalization import (
+    lower_all_char,
+    lower_only_alpha,
+    remove_non_alphabetic,
+    train_text_normalization,
+    upper_all_char,
+    upper_only_alpha,
+)
+from torch.utils.data.dataloader import DataLoader, default_collate
+
+
+class PromptASRDataset(torch.utils.data.Dataset):
+    """This is a dataset for Prompt ASR. It supports the following features:
+    1. Select a tuple of (text, pre_text, style_text) randomly from a
+    list of texts as supervisions.
+
+    """
+
+    def __init__(
+        self,
+        return_cuts: bool = False,
+        cut_transforms: List[Callable[[CutSet], CutSet]] = None,
+        input_transforms: List[Callable[[torch.Tensor], torch.Tensor]] = None,
+        input_strategy: BatchIO = PrecomputedFeatures(),
+        text_sampling_func: Optional[Callable[[List[str]], str]] = None,
+        rare_word_list: Optional[List[str]] = None,
+    ):
+        """
+        Icefall ASR IterableDataset constructor. See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py
+        for more details.
+
+        :param return_cuts: When ``True``, will additionally return a "cut" field in each batch with the Cut
+            objects used to create that batch.
+        :param cut_transforms: A list of transforms to be applied on each sampled batch,
+            before converting cuts to an input representation (audio/features).
+            Examples: cut concatenation, noise cuts mixing, etc.
+        :param input_transforms: A list of transforms to be applied on each sampled batch,
+            after the cuts are converted to audio/features.
+            Examples: normalization, SpecAugment, etc.
+        :param input_strategy: Converts cuts into a collated batch of audio/features.
+            By default, reads pre-computed features from disk.
+        :param text_sampling_func: Sampling a text as transcription from a list of texts.
+        """
+        super().__init__()
+        # Initialize the fields
+        self.return_cuts = return_cuts
+        self.cut_transforms = ifnone(cut_transforms, [])
+        self.input_transforms = ifnone(input_transforms, [])
+        self.input_strategy = input_strategy
+
+        # a text sampling function
+        self.text_sampling_func = text_sampling_func
+        self.rare_word_list = rare_word_list
+
+    def __getitem__(self, cuts: CutSet) -> Dict[str, Union[torch.Tensor, List[str]]]:
+        """
+        Return a new batch, with the batch size automatically determined using the constraints
+        of max_frames and max_cuts.
+        """
+        validate_for_asr(cuts)
+
+        # Sort the cuts by duration so that the first one determines the batch time dimensions.
+        cuts = cuts.sort_by_duration(ascending=False)
+
+        # Optional CutSet transforms - e.g. padding, or speed perturbation that adjusts
+        # the supervision boundaries.
+        for tnfm in self.cut_transforms:
+            cuts = tnfm(cuts)
+
+        # Sort the cuts again after transforms
+        cuts = cuts.sort_by_duration(ascending=False)
+
+        # Get a tensor with batched feature matrices, shape (B, T, F)
+        # Collation performs auto-padding, if necessary.
+        input_tpl = self.input_strategy(cuts)
+        if len(input_tpl) == 3:
+            # An input strategy with fault tolerant audio reading mode.
+            # "cuts" may be a subset of the original "cuts" variable,
+            # that only has cuts for which we succesfully read the audio.
+            inputs, _, cuts = input_tpl
+        else:
+            inputs, _ = input_tpl
+
+        # Get a dict of tensors that encode the positional information about supervisions
+        # in the batch of feature matrices. The tensors are named "sequence_idx",
+        # "start_frame/sample" and "num_frames/samples".
+        supervision_intervals = self.input_strategy.supervision_intervals(cuts)
+
+        # Apply all available transforms on the inputs, i.e. either audio or features.
+        # This could be feature extraction, global MVN, SpecAugment, etc.
+        segments = torch.stack(list(supervision_intervals.values()), dim=1)
+        for tnfm in self.input_transforms:
+            inputs = tnfm(inputs, supervision_segments=segments)
+
+        batch = {
+            "inputs": inputs,
+            "supervisions": default_collate(
+                [
+                    self.text_sampling_func(
+                        texts=supervision.texts,
+                        pre_texts=supervision.pre_texts,
+                        context_list=supervision.context_list
+                        if "context_list" in supervision.custom
+                        else None,
+                        rare_word_list=self.rare_word_list,
+                    )
+                    if self.text_sampling_func is not None
+                    else {
+                        "text": train_text_normalization(supervision.texts[0]),
+                        "pre_text": train_text_normalization(supervision.pre_texts[0]),
+                        "style_text": train_text_normalization(
+                            supervision.pre_texts[0]
+                        ),
+                        "transform_ids": 0,
+                    }
+                    for sequence_idx, cut in enumerate(cuts)
+                    for supervision in cut.supervisions
+                ]
+            ),
+        }
+        # Update the 'supervisions' field with sequence_idx and start/num frames/samples
+        batch["supervisions"].update(supervision_intervals)
+        if self.return_cuts:
+            batch["supervisions"]["cut"] = [
+                cut for cut in cuts for sup in cut.supervisions
+            ]
+
+        has_word_alignments = all(
+            s.alignment is not None and "word" in s.alignment
+            for c in cuts
+            for s in c.supervisions
+        )
+
+        return batch
+
+
+def validate_for_asr(cuts: CutSet) -> None:
+    validate(cuts)
+    tol = 2e-3  # 1ms
+    for cut in cuts:
+        for supervision in cut.supervisions:
+            assert supervision.start >= -tol, (
+                f"Supervisions starting before the cut are not supported for ASR"
+                f" (sup id: {supervision.id}, cut id: {cut.id})"
+            )
+
+            # Supervision start time is relative to Cut ...
+            # https://lhotse.readthedocs.io/en/v0.10_e/cuts.html
+            #
+            # 'supervision.end' is end of supervision inside the Cut
+            assert supervision.end <= cut.duration + tol, (
+                f"Supervisions ending after the cut "
+                f"are not supported for ASR"
+                f" (sup id: {supervision.id}, cut id: {cut.id})"
+            )
+
+
+def get_substring(s: str, min_len: int = 40, max_len: int = 250) -> str:
+    """A helper function that generates a random substring from a given string
+
+    Args:
+        s (str): Input string
+
+    Returns:
+        str: Returned substring
+    """
+    min_len = min(len(s), min_len)
+
+    start = random.randint(0, len(s) - min_len)
+    end = min(start + max_len, random.randint(start + min_len, len(s)))
+
+    return s[start:end]
+
+
+def triplet_text_sampling(
+    texts: List[str],
+    pre_texts: List[str],
+    context_list: Optional[str] = None,
+    rare_word_list: Optional[List[str]] = None,
+    transforms: Optional[List[Callable[[str], str]]] = None,
+    min_len_style: Optional[int] = 80,
+) -> Dict[str, str]:
+    """This function generates a triplet of
+    (pre_text, style_text, ref_text). The style of style_text and ref_text
+    should **always** match, whereas the style of pre_text is arbitrary.
+    Suppose we have 2 different transforms A,B, and the preceding text is
+    referred to as pre_text. The following three tuples are all valid:
+
+    (A(pre_text), A(style_text), A(ref_text))
+    (A(pre_text), B(style_text), B(ref_text))
+    (A(pre_text), A(style_text), A(ref_text))
+    (B(pre_text), B(style_text), B(ref_text))
+
+    If transforms is not given, the following pre-defined transforms
+    are available:
+    0: original (mixed-cased, with punc)
+    1: upper_only_alpha (upper-cased, no punc)
+
+    When the transform of text and pre_text match, we can use the whole
+    pre_text as the prompt text.
+
+    Args:
+        texts (List[str]):
+            A list of ref_texts whose first item is the ground truth
+            text from books.
+        pre_texts (List[str]):
+            A list of pre_texts, whose first item is the groundtruth
+            pre_text from books.
+        context_list: Optional[str] = None,
+            A list of biasing words separated by space
+        rare_word_list: Optional[str] = None,
+            A list of rare-words separated by space (used as distractors)
+        transforms (List[Callable[[str], str]]): A list of possible transforms to be applied
+
+    Returns:
+        A dictionary of ref_text, pre_text, style_text
+    """
+    assert len(texts) == len(pre_texts)
+    assert len(texts) == 2
+
+    # we assume the first item to be ground truth
+    gt_text = texts[0]
+    gt_pre_text = pre_texts[0]
+
+    if transforms is None:
+        transforms = [
+            lambda x: x,  # return it self
+            upper_only_alpha,
+            lower_only_alpha,
+            lower_all_char,
+        ]
+
+    sampling_weight = [
+        0.7,
+        0.3,
+        0.0,
+        0.0,
+    ]  # Mixed-punc should have the largest sampling prob
+
+    total_transforms = len(transforms)  # do not use the recognized trans
+
+    # Randomly sample transforms
+    i_text, i_pre_text = np.random.choice(total_transforms, 2, p=sampling_weight)
+
+    # get the normalized text and pre_text
+    text = transforms[i_text](gt_text)
+    pre_text = transforms[i_pre_text](gt_pre_text)
+
+    if i_text == i_pre_text:
+        style_text = get_substring(pre_text, min_len=min_len_style, max_len=150)
+    else:
+        # get the pre_text of same style as text
+        # For now, **don't** do transform to the style text, because we do it after the dataloader
+        style_text = gt_pre_text
+        # style_text = pre_texts[i_text] if i_text <= 1 else transforms[i_text-2](gt_pre_text)
+        style_text = get_substring(style_text, min_len=min_len_style, max_len=150)
+
+    return {
+        "text": train_text_normalization(text),
+        "pre_text": train_text_normalization(pre_text),
+        "style_text": train_text_normalization(style_text),
+        "transform_ids": i_text,
+    }
+
+
+def triplet_text_sampling_with_context_list(
+    texts: List[str],
+    pre_texts: List[str],
+    context_list: str,
+    rare_word_list: List[str],
+    transforms: Optional[List[Callable[[str], str]]] = None,
+    min_len_style: Optional[int] = 80,
+) -> Dict[str, str]:
+    """This function generates a triplet of
+    (pre_text, style_text, ref_text). The pre_text is either the preceding text
+    or a list of words (context words + distractors).
+    The style of style_text and ref_text should **always** match, whereas
+    the style of pre_text is arbitrary.
+    Suppose we have 2 different transforms A,B, and the preceding text is
+    referred to as pre_text. The following three tuples are all valid:
+
+    (A(pre_text), A(style_text), A(ref_text))
+    (A(pre_text), B(style_text), B(ref_text))
+    (A(pre_text), A(style_text), A(ref_text))
+    (B(pre_text), B(style_text), B(ref_text))
+
+    If transforms is not given, the following pre-defined transforms
+    are available:
+    0: original (mixed-cased, with punc)
+    1: upper_only_alpha (upper-cased, no punc)
+
+    When the transform of text and pre_text match, we can use the whole
+    pre_text as the prompt text.
+
+    Args:
+        texts (List[str]):
+            A list of ref_texts whose first item is the ground truth
+            text from books.
+        pre_texts (List[str]):
+            A list of pre_texts, whose first item is the groundtruth
+            pre_text from books.
+        context_list: Optional[str] = None,
+            A list of biasing words separated by space
+        rare_word_list: Optional[str] = None,
+            A list of rare-words separated by space (used as distractors)
+        transforms (List[Callable[[str], str]]): A list of possible transforms to be applied
+
+    Returns:
+        A dictionary of ref_text, pre_text, style_text
+    Returns:
+        str: A dictionary
+    """
+    # import pdb; pdb.set_trace()
+    assert len(texts) == len(pre_texts)
+    assert len(texts) == 2
+
+    if context_list is not None:
+        context_list = context_list.lower()
+
+    # we assume the first item to be ground truth
+    gt_text = texts[0]
+    gt_pre_text = pre_texts[0]
+
+    if transforms is None:
+        transforms = [
+            lambda x: x,  # return it self
+            upper_only_alpha,
+            lower_only_alpha,
+            lower_all_char,
+        ]
+
+    sampling_weight = [
+        0.7,
+        0.3,
+        0.0,
+        0.0,
+    ]  # Mixed-punc should have the largest sampling prob
+
+    total_transforms = len(transforms)  # do not use the recognized trans
+
+    # Select a transformation randomly
+    i_text, i_pre_text = np.random.choice(total_transforms, 2, p=sampling_weight)
+
+    # get the normalized text and pre_text
+    text = transforms[i_text](gt_text)
+    pre_text = get_pre_text_with_context_list2(
+        text=gt_text,
+        pre_text=gt_pre_text,
+        context_list=context_list,
+        rare_words_list=rare_word_list,
+    )
+    pre_text = transforms[i_pre_text](pre_text)
+
+    if i_text == i_pre_text:
+        style_text = get_substring(pre_text, min_len=min_len_style, max_len=150)
+    else:
+        # get the pre_text of same style as text
+        # For now, **don't** do transform to the style text
+        style_text = gt_pre_text
+        # style_text = pre_texts[i_text] if i_text <= 1 else transforms[i_text-2](gt_pre_text)
+        style_text = get_substring(style_text, min_len=min_len_style, max_len=150)
+
+    return {
+        "text": train_text_normalization(text),
+        "pre_text": train_text_normalization(pre_text),
+        "style_text": train_text_normalization(style_text),
+        "transform_ids": i_text,
+    }
+
+
+def get_pre_text_with_context_list(
+    text: str,
+    pre_text: str,
+    context_list: str,
+    rare_words_list: List[str] = None,
+) -> str:
+    # Always get the first one, which is the gt (mixed-cased trans), but with upper_only_alpha
+    # By a small proportion of time, use the substring of ref_text as pre_text
+
+    if context_list != "" and context_list is not None:
+        v = random.random()
+        if v < 0.5:
+            # correct + distractors
+            # sample distractors
+            num_distractors = random.randint(0, 50)
+            distractors = random.sample(rare_words_list, num_distractors)
+            # sample correct
+            correct = context_list.split()
+            i = random.randint(1, len(correct))
+            correct = random.sample(correct, i)
+            # combine correct and distractors
+            pre_text = distractors + correct
+            random.shuffle(pre_text)
+            pre_text = " ".join(pre_text)
+        elif v < 0.7:
+            splitted = text.split()
+            sampling_weights = [len(w) ** 1.2 for w in splitted]
+            sampling_weights = [p / sum(sampling_weights) for p in sampling_weights]
+            i = random.randint(1, min(len(splitted), 20))
+            splitted = list(np.random.choice(splitted, i, p=sampling_weights))
+            num_distractors = random.randint(0, 70)
+            distractors = random.sample(rare_words_list, num_distractors)
+            splitted += distractors
+            random.shuffle(splitted)  # shuffle the list
+            pre_text = " ".join(splitted)
+        else:
+            pre_text = pre_text
+    else:
+        v = random.random()
+        if v < 0.1:
+            splitted = text.split()
+            sampling_weights = [len(w) ** 1.2 for w in splitted]
+            sampling_weights = [p / sum(sampling_weights) for p in sampling_weights]
+            i = random.randint(1, min(len(splitted), 20))
+            splitted = list(np.random.choice(splitted, i, p=sampling_weights))
+            pre_text = " ".join(splitted)
+            num_distractors = random.randint(0, 70)
+            distractors = random.sample(rare_words_list, num_distractors)
+            splitted += distractors
+            random.shuffle(splitted)  # shuffle the list
+        elif v < 0.2:
+            # full distractors
+            num_distractors = random.randint(5, 100)
+            distractors = random.sample(rare_words_list, num_distractors)
+            pre_text = " ".join(distractors)
+
+        elif v < 0.3:
+            pre_text = get_substring(text, min_len=15, max_len=150)
+        else:
+            pre_text = pre_text
+
+    return pre_text
+
+
+def get_pre_text_with_context_list2(
+    text: str,
+    pre_text: str,
+    context_list: str,
+    rare_words_list: List[str] = None,
+) -> str:
+    # Get the pre_text, either the ground truth preceding text or
+    # a list of words consisting of biasing words and distrators
+    # By a small proportion of time, use the substring of ref_text as pre_text
+
+    if context_list != "" and context_list is not None:
+        v = random.random()
+        if v < 0.4:
+            # sample distractors
+            num_distractors = random.randint(50, 100)
+            distractors = random.sample(rare_words_list, num_distractors)
+            # sample correct
+            correct = context_list.split()
+            i = random.randint(1, len(correct))
+            correct = random.sample(correct, i)
+            # combine correct and distractors
+            pre_text = distractors + correct
+            random.shuffle(pre_text)
+            pre_text = " ".join(pre_text)
+        elif v < 0.55:
+            splitted = text.split()
+            sampling_weights = [
+                len(w) ** 1.2 for w in splitted
+            ]  # longer words with higher weights
+            sampling_weights = [p / sum(sampling_weights) for p in sampling_weights]
+            i = random.randint(1, min(len(splitted), 20))
+            splitted = list(np.random.choice(splitted, i, p=sampling_weights))
+            num_distractors = random.randint(50, 100)
+            distractors = random.sample(rare_words_list, num_distractors)
+            splitted += distractors
+            random.shuffle(splitted)  # shuffle the list
+            pre_text = " ".join(splitted)
+        else:
+            pre_text = pre_text
+    else:
+        v = random.random()
+        if v < 0.3:
+            splitted = text.split()
+            sampling_weights = [len(w) ** 1.2 for w in splitted]
+            sampling_weights = [p / sum(sampling_weights) for p in sampling_weights]
+            i = random.randint(1, min(len(splitted), 20))
+            splitted = list(np.random.choice(splitted, i, p=sampling_weights))
+            pre_text = " ".join(splitted)
+            num_distractors = random.randint(50, 100)
+            distractors = random.sample(rare_words_list, num_distractors)
+            splitted += distractors
+            random.shuffle(splitted)  # shuffle the list
+        elif v < 0.4:
+            # full distractors
+            num_distractors = random.randint(5, 100)
+            distractors = random.sample(rare_words_list, num_distractors)
+            pre_text = " ".join(distractors)
+        elif v < 0.6:
+            pre_text = get_substring(text, min_len=15, max_len=150)
+        else:
+            pre_text = pre_text
+
+    return pre_text
+
+
+def naive_triplet_text_sampling(
+    texts: List[str],
+    pre_texts: List[str],
+    context_list: str = None,
+    rare_word_list: List[str] = None,
+    min_len_style: Optional[int] = 120,
+):
+    # The most simplest text sampling function, used only for
+    # evaluation, use a fixed sentence as the style text
+
+    return {
+        "text": train_text_normalization(texts[0]),
+        "pre_text": train_text_normalization(pre_texts[0]),
+        "style_text": "Mixed-case English transcription, with punctuation. Actually, it is fully not related. What do you think?",
+        "transform_ids": 0,
+    }
+
+
+def random_shuffle_subset(
+    data: List[str],
+    p: float = 0.2,
+    p_mask: float = 0.05,
+) -> List[str]:
+    """
+    Randomly shuffle the subset by probability `p`, which means that p% of the samples
+    in the original batch are shuffled, the others are kept in the original order.
+
+    With a probability of `p_mask`, replace the original string with an empty string.
+
+    """
+
+    num_to_shuffle = int(len(data) * p)
+    id_to_shuffle = np.random.choice(len(data), num_to_shuffle, replace=False)
+    item_to_shuffle = [data[id] for id in id_to_shuffle]
+    random.shuffle(item_to_shuffle)
+
+    for id, item in zip(id_to_shuffle, item_to_shuffle):
+        data[id] = item
+
+    # Randomly mask a proportion of the data to empty string
+    if p_mask > 0:
+        for i in range(len(data)):
+            if random.random() < p_mask:
+                data[i] = ""
+
+    return data
+
+
+if __name__ == "__main__":
+    texts = [
+        "AA, BB, cC, dD!",
+        "AA BB CC DD",
+    ]
+
+    pre_texts = [
+        "EE, Ff, Gg? EE, Ff, Gg? EE, Ff, Gg? EE, Ff, Gg?",
+        "EE FF GG EE FF GG EE FF GG EE FF GG EE FF GG",
+    ]
+    for i in range(10):
+        print(f"Run: {i}")
+        print(triplet_text_sampling(texts, pre_texts))
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/decode_baseline.py b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_baseline.py
new file mode 100644
index 000000000..6a3bab3c8
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_baseline.py
@@ -0,0 +1,791 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) modified beam search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+"""
+
+
+import argparse
+import logging
+import math
+import warnings
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriHeavyAsrDataModule
+from beam_search import greedy_search, greedy_search_batch, modified_beam_search
+from ls_text_normalization import word_normalization
+from text_normalization import (
+    ref_text_normalization,
+    remove_non_alphabetic,
+    upper_only_alpha,
+)
+from train_baseline import add_model_arguments, get_params, get_transducer_model
+from utils import write_error_stats
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, setup_logger, store_transcripts, str2bool
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion
+          - modified_beam_search_LODR
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--post-normalization",
+        type=str2bool,
+        default=True,
+        help="Normalized the recognition results by uppercasing and removing non-alphabetic symbols. ",
+    )
+
+    parser.add_argument(
+        "--long-audio-recog",
+        type=str2bool,
+        default=False,
+    )
+
+    parser.add_argument(
+        "--use-ls-test-set",
+        type=str2bool,
+        default=False,
+        help="Use librispeech test set for evaluation.",
+    )
+
+    parser.add_argument(
+        "--compute-CER",
+        type=str2bool,
+        default=True,
+        help="Reports CER. By default, only reports WER",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural net LM for shallow fusion. Only used when `--use-shallow-fusion`
+        set to true.
+      ngram_lm:
+        A ngram lm. Used in LODR decoding.
+      ngram_lm_scale:
+        The scale of the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    texts = batch["supervisions"]["text"]
+    batch_size = feature.size(0)
+
+    # Get the transducer encoder output
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        encoder_out, encoder_out_lens = model.encode_audio(
+            feature=feature,
+            feature_lens=feature_lens,
+        )
+
+    hyps = []
+
+    if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        if not params.use_ls_test_set:
+            book_names = [
+                cut.text_path.split("/")[-2] for cut in batch["supervisions"]["cut"]
+            ]
+        else:
+            book_names = ["" for _ in cut_ids]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, book_name, hyp_words, ref_text in zip(
+                cut_ids, book_names, hyps, texts
+            ):
+                ref_text = ref_text_normalization(ref_text)
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+            # if not params.use_ls_test_set:
+            #     results[name + " " + book_name].extend(this_batch)
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+    biasing_words: List[str] = None,
+):
+    test_set_wers = dict()
+    test_set_cers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f,
+                f"{test_set_name}-{key}",
+                results,
+                enable_log=True,
+                biasing_words=biasing_words,
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+        if params.compute_CER:
+            # Write CER statistics
+            recog_path = (
+                params.res_dir / f"recogs-{test_set_name}-char-{params.suffix}.txt"
+            )
+            store_transcripts(filename=recog_path, texts=results, char_level=True)
+            errs_filename = (
+                params.res_dir / f"errs-CER-{test_set_name}-{params.suffix}.txt"
+            )
+            with open(errs_filename, "w") as f:
+                cer = write_error_stats(
+                    f,
+                    f"{test_set_name}-{key}",
+                    results,
+                    enable_log=True,
+                    compute_CER=params.compute_CER,
+                )
+                test_set_cers[key] = cer
+
+            logging.info("Wrote detailed CER stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+    if params.compute_CER:
+        test_set_cers = sorted(test_set_cers.items(), key=lambda x: x[1])
+        errs_info = params.res_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt"
+        with open(errs_info, "w") as f:
+            print("settings\tcER", file=f)
+            for key, val in test_set_cers:
+                print("{}\t{}".format(key, val), file=f)
+
+        s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+        note = "\tbest for {}".format(test_set_name)
+        for key, val in test_set_cers:
+            s += "{} CER\t{}{}\n".format(key, val, note)
+            note = ""
+        logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "modified_beam_search",
+    )
+
+    if params.long_audio_recog:
+        params.res_dir = params.exp_dir / (params.decoding_method + "long_audio")
+    else:
+        params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if "ngram" in params.decoding_method:
+        params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to(device)
+    model.eval()
+
+    LM = None
+
+    decoding_graph = None
+    word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    test_clean_cuts = libriheavy.test_clean_cuts()
+    test_other_cuts = libriheavy.test_other_cuts()
+    ls_test_clean_cuts = libriheavy.librispeech_test_clean_cuts()
+    ls_test_other_cuts = libriheavy.librispeech_test_other_cuts()
+    long_audio_cuts = libriheavy.long_audio_cuts()
+
+    test_clean_dl = libriheavy.valid_dataloaders(
+        test_clean_cuts,
+    )
+    test_other_dl = libriheavy.valid_dataloaders(
+        test_other_cuts,
+    )
+    ls_test_clean_dl = libriheavy.test_dataloaders(ls_test_clean_cuts)
+    ls_test_other_dl = libriheavy.test_dataloaders(ls_test_other_cuts)
+    long_audio_dl = libriheavy.valid_dataloaders(
+        long_audio_cuts,
+    )
+
+    if params.use_ls_test_set:
+        test_sets = ["ls-test-clean", "ls-test-other"]
+        test_dl = [ls_test_clean_dl, ls_test_other_dl]
+    else:
+        test_sets = ["test-clean", "test-other"]
+        test_dl = [test_clean_dl, test_other_dl]
+
+    if params.long_audio_recog:
+        test_sets = ["long-audio"]
+        test_dl = [long_audio_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        if params.use_ls_test_set:
+            f = open(
+                "data/context_biasing/LibriSpeechBiasingLists/all_rare_words.txt", "r"
+            )
+            biasing_words = f.read().strip().split()
+            f.close()
+        else:
+            biasing_words = None
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+        if params.post_normalization:
+            if "-post-normalization" not in params.suffix:
+                params.suffix += "-post-normalization"
+
+            new_res = {}
+            for k in results_dict:
+                new_ans = []
+                for item in results_dict[k]:
+                    id, ref, hyp = item
+                    if params.use_ls_test_set:
+                        hyp = (
+                            " ".join(hyp).replace("-", " ").split()
+                        )  # handle the hypens
+                        hyp = upper_only_alpha(" ".join(hyp)).split()
+                        hyp = [word_normalization(w.upper()) for w in hyp]
+                        hyp = " ".join(hyp).split()
+                        hyp = [w for w in hyp if w != ""]
+                        ref = upper_only_alpha(" ".join(ref)).split()
+                    else:
+                        hyp = upper_only_alpha(" ".join(hyp)).split()
+                        ref = upper_only_alpha(" ".join(ref)).split()
+                    new_ans.append((id, ref, hyp))
+                new_res[k] = new_ans
+
+            save_results(
+                params=params,
+                test_set_name=test_set,
+                results_dict=new_res,
+                biasing_words=biasing_words,
+            )
+
+            if params.suffix.endswith("-post-normalization"):
+                params.suffix = params.suffix.replace("-post-normalization", "")
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert.py b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert.py
new file mode 100755
index 000000000..e71999b0a
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert.py
@@ -0,0 +1,1025 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./zipformer_prompt_asr/decode_bert.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --max-duration 1000 \
+    --decoding-method greedy_search \
+    --text-encoder-type BERT \
+    --memory-layer 0 \
+    --use-pre-text True \
+    --use-style-prompt True \
+    --max-prompt-lens 1000 \
+    --style-text-transform mixed-punc \
+    --pre-text-transform mixed-punc \
+    --compute-CER 0
+
+
+(2) modified beam search
+./zipformer_prompt_asr/decode_bert.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --max-duration 1000 \
+    --decoding-method modified_beam_search \
+    --beam-size 4 \
+    --text-encoder-type BERT \
+    --memory-layer 0 \
+    --use-pre-text True \
+    --use-style-prompt True \
+    --max-prompt-lens 1000 \
+    --style-text-transform mixed-punc \
+    --pre-text-transform mixed-punc \
+    --compute-CER 0
+
+(3) Decode LibriSpeech
+
+./zipformer_prompt_asr/decode_bert.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --max-duration 1000 \
+    --decoding-method modified_beam_search \
+    --use-ls-test-set True \
+    --beam-size 4 \
+    --text-encoder-type BERT \
+    --memory-layer 0 \
+    --use-pre-text True \
+    --use-style-prompt True \
+    --max-prompt-lens 1000 \
+    --style-text-transform mixed-punc \
+    --pre-text-transform mixed-punc \
+    --compute-CER 0
+
+(4) Decode LibriSpeech + biasing list
+
+biasing_list=100 # could also be 0
+
+./zipformer_prompt_asr/decode_bert.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --max-duration 1000 \
+    --decoding-method modified_beam_search \
+    --beam-size 4 \
+    --use-ls-test-set True \
+    --use-ls-context-list True \
+    --biasing-level utterance \
+    --ls-distractors $biasing_list \
+    --post-normalization True \
+    --text-encoder-type BERT \
+    --max-prompt-lens 1000 \
+    --style-text-transform mixed-punc \
+    --pre-text-transform mixed-punc
+
+
+"""
+
+
+import argparse
+import logging
+import math
+import warnings
+from collections import defaultdict
+from pathlib import Path
+from typing import Callable, Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriHeavyAsrDataModule
+from beam_search import greedy_search, greedy_search_batch, modified_beam_search
+from dataset import naive_triplet_text_sampling, random_shuffle_subset
+from ls_text_normalization import word_normalization
+from text_normalization import (
+    _apply_style_transform,
+    lower_all_char,
+    lower_only_alpha,
+    ref_text_normalization,
+    remove_non_alphabetic,
+    train_text_normalization,
+    upper_all_char,
+    upper_only_alpha,
+)
+from train_bert_encoder import (
+    _encode_texts_as_bytes_with_tokenizer,
+    add_model_arguments,
+    get_params,
+    get_tokenizer,
+    get_transducer_model,
+)
+from transformers import BertModel, BertTokenizer
+from utils import brian_biasing_list, get_facebook_biasing_list, write_error_stats
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, setup_logger, store_transcripts, str2bool
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion
+          - modified_beam_search_LODR
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--use-pre-text",
+        type=str2bool,
+        default=True,
+        help="Use pre-text is available during decoding",
+    )
+
+    parser.add_argument(
+        "--use-style-prompt",
+        type=str2bool,
+        default=True,
+        help="Use style prompt when evaluation",
+    )
+
+    parser.add_argument(
+        "--max-prompt-lens",
+        type=int,
+        default=1000,
+    )
+
+    parser.add_argument(
+        "--post-normalization",
+        type=str2bool,
+        default=True,
+        help="Normalized the recognition results by uppercasing and removing non-alphabetic symbols. ",
+    )
+
+    parser.add_argument(
+        "--compute-CER",
+        type=str2bool,
+        default=False,
+        help="Reports CER. By default, only reports WER",
+    )
+
+    parser.add_argument(
+        "--style-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of style prompt, i.e style_text",
+    )
+
+    parser.add_argument(
+        "--pre-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of content prompt, i.e pre_text",
+    )
+
+    parser.add_argument(
+        "--use-ls-test-set",
+        type=str2bool,
+        default=False,
+        help="Use librispeech test set for evaluation.",
+    )
+
+    parser.add_argument(
+        "--use-ls-context-list",
+        type=str2bool,
+        default=False,
+        help="If use a fixed context list for LibriSpeech decoding",
+    )
+
+    parser.add_argument(
+        "--biasing-level",
+        type=str,
+        default="utterance",
+        choices=["utterance", "Book", "Chapter"],
+    )
+
+    parser.add_argument(
+        "--ls-distractors",
+        type=int,
+        default=0,
+        help="The number of distractors into context list for LibriSpeech decoding",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    tokenizer: spm.SentencePieceProcessor,
+    batch: dict,
+    biasing_dict: dict = None,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      tokenizer:
+        Tokenizer for the text encoder
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      biasing_dict:
+        A dictionary in the form `{cut_id: :w1 w2"}` that contains a list
+        of biasing words (separated with space)
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural net LM for shallow fusion. Only used when `--use-shallow-fusion`
+        set to true.
+      ngram_lm:
+        A ngram lm. Used in LODR decoding.
+      ngram_lm_scale:
+        The scale of the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    cuts = batch["supervisions"]["cut"]
+    cut_ids = [c.supervisions[0].id for c in cuts]
+    batch_size = feature.size(0)
+
+    if "pre_text" in batch["supervisions"] and params.use_pre_text:
+        pre_texts = batch["supervisions"]["pre_text"]
+        pre_texts = [train_text_normalization(t) for t in pre_texts]
+    else:
+        pre_texts = ["" for _ in range(batch_size)]
+
+    # get the librispeech biasing data
+    if params.use_pre_text and (params.use_ls_context_list and params.use_ls_test_set):
+        if params.biasing_level == "utterance":
+            pre_texts = [biasing_dict[id] for id in cut_ids]
+        elif params.biasing_level == "Chapter":
+            chapter_ids = [c.split("-")[1] for c in cut_ids]
+            pre_texts = [biasing_dict[id] for id in chapter_ids]
+        elif params.biasing_level == "Book":
+            chapter_ids = [c.split("-")[1] for c in cut_ids]
+            pre_texts = [biasing_dict[id] for id in chapter_ids]
+        else:
+            raise ValueError(f"Unseen biasing level: {params.biasing_level}")
+        if params.pre_text_transform == "mixed-punc":
+            pre_texts = [t.lower() for t in pre_texts]
+
+    # get style_text
+    if params.use_style_prompt:
+        fixed_sentence = "Mixed-case English transcription, with punctuation. Actually, it's fully not related."
+        style_texts = batch["supervisions"].get(
+            "style_text", [fixed_sentence for _ in range(batch_size)]
+        )
+        style_texts = [train_text_normalization(t) for t in style_texts]
+    else:
+        style_texts = ["" for _ in range(batch_size)]  # use empty string
+
+    # Get the text embedding
+    if params.use_pre_text or params.use_style_prompt:
+        # apply style transform to the pre_text and style_text
+        pre_texts = _apply_style_transform(pre_texts, params.pre_text_transform)
+        if not params.use_ls_context_list:
+            pre_texts = [t[-params.max_prompt_lens :] for t in pre_texts]
+
+        if params.use_style_prompt:
+            style_texts = _apply_style_transform(
+                style_texts, params.style_text_transform
+            )
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+
+            # Use tokenizer to prepare input for text encoder
+            encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer(
+                pre_texts=pre_texts,
+                style_texts=style_texts,
+                tokenizer=tokenizer,
+                device=device,
+                no_limit=True,
+            )
+            logging.info(
+                f"Shape of the encoded prompts: {encoded_inputs['input_ids'].shape}"
+            )
+
+            memory, memory_key_padding_mask = model.encode_text(
+                encoded_inputs=encoded_inputs,
+                style_lens=style_lens,
+            )  # (T,B,C)
+    else:
+        memory = None
+        memory_key_padding_mask = None
+
+    # Get the transducer encoder output
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        encoder_out, encoder_out_lens = model.encode_audio(
+            feature=feature,
+            feature_lens=feature_lens,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )
+
+    hyps = []
+
+    if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    tokenizer: spm.SentencePieceProcessor,
+    biasing_dict: Dict = None,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      tokenizer:
+        Tokenizer for the text encoder
+      biasing_dict:
+        A dictionary in the form `{cut_id: :w1 w2"}` that contains a list
+        of biasing words (separated with space)
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"][
+            "text"
+        ]  # By default, this should be in mixed-punc format
+
+        # the style of ref_text should match style_text
+        texts = _apply_style_transform(texts, params.style_text_transform)
+        if params.use_style_prompt:
+            texts = _apply_style_transform(texts, params.style_text_transform)
+
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        if not params.use_ls_test_set:
+            try:
+                book_names = [
+                    cut.text_path.split("/")[-2] for cut in batch["supervisions"]["cut"]
+                ]
+            except AttributeError:
+                book_names = [
+                    cut.id.split("/")[0] for cut in batch["supervisions"]["cut"]
+                ]
+        else:
+            book_names = ["" for _ in cut_ids]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            tokenizer=tokenizer,
+            biasing_dict=biasing_dict,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, book_name, hyp_words, ref_text in zip(
+                cut_ids, book_names, hyps, texts
+            ):
+                ref_text = ref_text_normalization(
+                    ref_text
+                )  # remove full-width symbols & some book marks
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+    biasing_words: List[str] = None,
+):
+    test_set_wers = dict()
+    test_set_cers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+        if params.compute_CER:
+            # Write CER statistics
+            recog_path = (
+                params.res_dir / f"recogs-{test_set_name}-char-{params.suffix}.txt"
+            )
+            store_transcripts(filename=recog_path, texts=results, char_level=True)
+            errs_filename = (
+                params.res_dir / f"errs-CER-{test_set_name}-{params.suffix}.txt"
+            )
+            with open(errs_filename, "w") as f:
+                cer = write_error_stats(
+                    f,
+                    f"{test_set_name}-{key}",
+                    results,
+                    enable_log=True,
+                    compute_CER=params.compute_CER,
+                )
+                test_set_cers[key] = cer
+
+            logging.info("Wrote detailed CER stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+    if params.compute_CER:
+        test_set_cers = sorted(test_set_cers.items(), key=lambda x: x[1])
+        errs_info = params.res_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt"
+        with open(errs_info, "w") as f:
+            print("settings\tCER", file=f)
+            for key, val in test_set_cers:
+                print("{}\t{}".format(key, val), file=f)
+
+        s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+        note = "\tbest for {}".format(test_set_name)
+        for key, val in test_set_cers:
+            s += "{} CER\t{}{}\n".format(key, val, note)
+            note = ""
+        logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "modified_beam_search",
+    )
+
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+        params.suffix += f"-chunk-{params.chunk_size}"
+        params.suffix += f"-left-context-{params.left_context_frames}"
+
+    if "beam_search" in params.decoding_method:
+        params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_pre_text:
+        params.suffix += (
+            f"-pre-text-{params.pre_text_transform}-len-{params.max_prompt_lens}"
+        )
+
+    if params.use_style_prompt:
+        params.suffix += f"-style-prompt-{params.style_text_transform}"
+
+    if params.use_ls_context_list:
+        assert (
+            params.use_pre_text
+        ), "Must set --use-pre-text to True if using context list"
+        params.suffix += f"-use-{params.biasing_level}-level-ls-context-list"
+        if params.biasing_level == "utterance" and params.ls_distractors:
+            params.suffix += f"-ls-context-distractors-{params.ls_distractors}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    tokenizer = get_tokenizer(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to(device)
+    model.eval()
+
+    LM = None
+
+    decoding_graph = None
+    word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    test_clean_cuts = libriheavy.test_clean_cuts()
+    test_other_cuts = libriheavy.test_other_cuts()
+    ls_test_clean_cuts = libriheavy.librispeech_test_clean_cuts()
+    ls_test_other_cuts = libriheavy.librispeech_test_other_cuts()
+
+    test_clean_dl = libriheavy.valid_dataloaders(
+        test_clean_cuts, text_sampling_func=naive_triplet_text_sampling
+    )
+    test_other_dl = libriheavy.valid_dataloaders(
+        test_other_cuts, text_sampling_func=naive_triplet_text_sampling
+    )
+    ls_test_clean_dl = libriheavy.test_dataloaders(ls_test_clean_cuts)
+    ls_test_other_dl = libriheavy.test_dataloaders(ls_test_other_cuts)
+
+    if params.use_ls_test_set:
+        test_sets = ["ls-test-clean", "ls-test-other"]
+        test_dl = [ls_test_clean_dl, ls_test_other_dl]
+    else:
+        test_sets = ["test-clean", "test-other"]
+        test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        biasing_dict = None
+        if params.use_ls_context_list:
+            if test_set == "ls-test-clean":
+                biasing_dict = get_facebook_biasing_list(
+                    test_set="test-clean",
+                    num_distractors=params.ls_distractors,
+                )
+            elif test_set == "ls-test-other":
+                biasing_dict = get_facebook_biasing_list(
+                    test_set="test-other",
+                    num_distractors=params.ls_distractors,
+                )
+
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            tokenizer=tokenizer,
+            biasing_dict=biasing_dict,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+        if params.post_normalization:
+            if "-post-normalization" not in params.suffix:
+                params.suffix += "-post-normalization"
+
+            new_res = {}
+            for k in results_dict:
+                new_ans = []
+                for item in results_dict[k]:
+                    id, ref, hyp = item
+                    if params.use_ls_test_set:
+                        hyp = (
+                            " ".join(hyp).replace("-", " ").split()
+                        )  # handle the hypens
+                        hyp = upper_only_alpha(" ".join(hyp)).split()
+                        hyp = [word_normalization(w.upper()) for w in hyp]
+                        hyp = " ".join(hyp).split()
+                        hyp = [w for w in hyp if w != ""]
+                        ref = upper_only_alpha(" ".join(ref)).split()
+                    else:
+                        hyp = upper_only_alpha(" ".join(hyp)).split()
+                        ref = upper_only_alpha(" ".join(ref)).split()
+                    new_ans.append((id, ref, hyp))
+                new_res[k] = new_ans
+
+            save_results(
+                params=params,
+                test_set_name=test_set,
+                results_dict=new_res,
+            )
+
+            if params.suffix.endswith("-post-normalization"):
+                params.suffix = params.suffix.replace("-post-normalization", "")
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert_with_style_save_decoding_mp.py b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert_with_style_save_decoding_mp.py
new file mode 100755
index 000000000..4559ebb6d
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert_with_style_save_decoding_mp.py
@@ -0,0 +1,963 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao,
+#                                                 Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) modified beam search
+./pruned_transducer_stateless7/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless7/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+"""
+
+
+import argparse
+import logging
+import math
+import warnings
+from collections import defaultdict
+from pathlib import Path
+from typing import Callable, Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriHeavyAsrDataModule
+from beam_search import (
+    greedy_search,
+    greedy_search_batch,
+    greedy_search_batch_with_context,
+    greedy_search_with_context,
+    modified_beam_search,
+)
+from dataset import naive_triplet_text_sampling, random_shuffle_subset
+from lhotse import load_manifest_lazy
+from text_normalization import (
+    lower_all_char,
+    lower_only_alpha,
+    ref_text_normalization,
+    remove_non_alphabetic,
+    train_text_normalization,
+    upper_all_char,
+    upper_only_alpha,
+)
+from train_bert_encoder_with_style import (
+    _encode_texts_as_bytes_with_tokenizer,
+    add_model_arguments,
+    get_params,
+    get_tokenizer,
+    get_transducer_model,
+)
+from transformers import BertModel, BertTokenizer
+from utils import get_facebook_biasing_list
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--log-dir",
+        type=str,
+        required=True,
+        help="Where to store the logs",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+          - modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion
+          - modified_beam_search_LODR
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=20.0,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=64,
+        help="""Used only when --decoding-method is
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--input-manifest",
+        type=str,
+        required=True,
+        help="The input manifest to be decoded",
+    )
+
+    parser.add_argument(
+        "--output-manifest",
+        type=str,
+        required=True,
+        help="Where to store the output manifest (directory)",
+    )
+
+    parser.add_argument(
+        "--use-pre-text",
+        type=str2bool,
+        default=True,
+        help="Use pre-text is available during decoding",
+    )
+
+    parser.add_argument(
+        "--use-style-prompt",
+        type=str2bool,
+        default=True,
+        help="Use style prompt when evaluation",
+    )
+
+    parser.add_argument(
+        "--use-context-embedding",
+        type=str2bool,
+        default=False,
+        help="Use context fuser when evaluation",
+    )
+
+    parser.add_argument(
+        "--post-normalization",
+        type=str2bool,
+        default=True,
+        help="Normalized the recognition results by uppercasing and removing non-alphabetic symbols. ",
+    )
+
+    parser.add_argument(
+        "--compute-CER",
+        type=str2bool,
+        default=True,
+        help="Reports CER. By default, only reports WER",
+    )
+
+    parser.add_argument(
+        "--style-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of style prompt, i.e style_text",
+    )
+
+    parser.add_argument(
+        "--pre-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of content prompt, i.e pre_text",
+    )
+
+    parser.add_argument(
+        "--use-ls-test-set",
+        type=str2bool,
+        default=False,
+        help="Use librispeech test set for evaluation.",
+    )
+
+    parser.add_argument(
+        "--use-ls-context-list",
+        type=str2bool,
+        default=False,
+        help="If use a fixed context list for LibriSpeech decoding",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def _apply_style_transform(text: List[str], transform: str) -> List[str]:
+    """Apply transform to a list of text. By default, the text are in
+    ground truth format, i.e mixed-punc.
+
+    Args:
+        text (List[str]): Input text string
+        transform (str): Transform to be applied
+
+    Returns:
+        List[str]: _description_
+    """
+    if transform == "mixed-punc":
+        return text
+    elif transform == "upper-no-punc":
+        return [upper_only_alpha(s) for s in text]
+    elif transform == "lower-no-punc":
+        return [lower_only_alpha(s) for s in text]
+    elif transform == "lower-punc":
+        return [lower_all_char(s) for s in text]
+    else:
+        raise NotImplementedError(f"Unseen transform: {transform}")
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    tokenizer,
+    batch: dict,
+    biasing_dict: dict = None,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural net LM for shallow fusion. Only used when `--use-shallow-fusion`
+        set to true.
+      ngram_lm:
+        A ngram lm. Used in LODR decoding.
+      ngram_lm_scale:
+        The scale of the ngram language model.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    cuts = batch["supervisions"]["cut"]
+    cut_ids = [c.supervisions[0].id for c in cuts]
+    batch_size = feature.size(0)
+
+    # get pre_text
+    if "pre_text" in batch["supervisions"] and params.use_pre_text:
+        pre_texts = batch["supervisions"][
+            "text"
+        ]  # use the ground truth ref text as pre_text
+        pre_texts = [train_text_normalization(t) for t in pre_texts]
+    else:
+        pre_texts = ["" for _ in range(batch_size)]
+
+    if params.use_ls_context_list:
+        pre_texts = [biasing_dict[id] for id in cut_ids]
+
+    # get style_text
+    if params.use_style_prompt:
+        fixed_sentence = "Mixed-case English transcription, with punctuation. Actually, it's fully not related."
+        style_texts = batch["supervisions"].get(
+            "style_text", [fixed_sentence for _ in range(batch_size)]
+        )
+        style_texts = [train_text_normalization(t) for t in style_texts]
+    else:
+        style_texts = ["" for _ in range(batch_size)]  # use empty string
+
+    # Get the text embedding input
+    if params.use_pre_text or params.use_style_prompt:
+
+        # apply style transform to the pre_text and style_text
+        pre_texts = _apply_style_transform(pre_texts, params.pre_text_transform)
+        # pre_texts = random_shuffle_subset(pre_texts, p=1.0, p_mask=0.0)
+        if params.use_style_prompt:
+            style_texts = _apply_style_transform(
+                style_texts, params.style_text_transform
+            )
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+
+            # Use tokenizer to prepare input for text encoder
+            encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer(
+                pre_texts=pre_texts,
+                style_texts=style_texts,
+                tokenizer=tokenizer,
+                device=device,
+            )
+
+            memory, memory_key_padding_mask = model.encode_text(
+                encoded_inputs=encoded_inputs,
+                style_lens=style_lens,
+            )  # (T,B,C)
+    else:
+        memory = None
+        memory_key_padding_mask = None
+
+    # Get the transducer encoder output
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        encoder_out, encoder_out_lens = model.encode_audio(
+            feature=feature,
+            feature_lens=feature_lens,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )
+
+    hyps = []
+
+    if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
+        if memory is None or not params.use_context_embedding:
+            hyp_tokens = greedy_search_batch(
+                model=model,
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+            )
+        else:
+            memory = memory.permute(1, 0, 2)  # (T,N,C) -> (N,T,C)
+            context = model.context_fuser(
+                memory, padding_mask=memory_key_padding_mask
+            )  # (N,C)
+            context = model.joiner.context_proj(context)  # (N,C)
+            hyp_tokens = greedy_search_batch_with_context(
+                model=model,
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+                context=context,
+            )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                if memory is None or not params.use_context_embedding:
+                    hyp = greedy_search(
+                        model=model,
+                        encoder_out=encoder_out_i,
+                        max_sym_per_frame=params.max_sym_per_frame,
+                    )
+                else:
+                    cur_context = context[i : i + 1, :]
+                    hyp = greedy_search_with_context(
+                        model=model,
+                        encoder_out=encoder_out_i,
+                        context=cur_context,
+                        max_sym_per_frame=params.max_sym_per_frame,
+                    )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    tokenizer,
+    biasing_dict: Dict = None,
+    word_table: Optional[k2.SymbolTable] = None,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      word_table:
+        The word symbol table.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
+      LM:
+        A neural network LM, used during shallow fusion
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 40
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"][
+            "text"
+        ]  # By default, this should be in mixed-punc format
+
+        # the style of ref_text should match style_text
+        texts = _apply_style_transform(texts, params.style_text_transform)
+        if params.use_style_prompt:
+            texts = _apply_style_transform(texts, params.style_text_transform)
+
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            tokenizer=tokenizer,
+            biasing_dict=biasing_dict,
+            decoding_graph=decoding_graph,
+            word_table=word_table,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_text = ref_text_normalization(
+                    ref_text
+                )  # remove full-width symbols & some book marks
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    test_set_cers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+        if params.compute_CER:
+            # Write CER statistics
+            recog_path = (
+                params.res_dir / f"recogs-{test_set_name}-char-{params.suffix}.txt"
+            )
+            store_transcripts(filename=recog_path, texts=results, char_level=True)
+            errs_filename = (
+                params.res_dir / f"errs-CER-{test_set_name}-{params.suffix}.txt"
+            )
+            with open(errs_filename, "w") as f:
+                cer = write_error_stats(
+                    f,
+                    f"{test_set_name}-{key}",
+                    results,
+                    enable_log=True,
+                    compute_CER=params.compute_CER,
+                )
+                test_set_cers[key] = cer
+
+            logging.info("Wrote detailed CER stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+    if params.compute_CER:
+        test_set_cers = sorted(test_set_cers.items(), key=lambda x: x[1])
+        errs_info = params.res_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt"
+        with open(errs_info, "w") as f:
+            print("settings\tCER", file=f)
+            for key, val in test_set_cers:
+                print("{}\t{}".format(key, val), file=f)
+
+        s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+        note = "\tbest for {}".format(test_set_name)
+        for key, val in test_set_cers:
+            s += "{} CER\t{}{}\n".format(key, val, note)
+            note = ""
+        logging.info(s)
+
+
+def add_decoding_result_to_manifest(
+    in_manifest,
+    out_manifest: str,
+    results_dict: Dict,
+):
+    # write the decoding results with prompt to the manifest as an
+    # extra ref text
+    new_ans = {}
+    for key, value in results_dict.items():
+        for items in value:
+            id, ref, hyp = items
+            new_ans[id] = " ".join(hyp)
+
+    def _add_decoding(c):
+        key = c.supervisions[0].id
+        c.supervisions[0].texts.append(new_ans[key])
+        return c
+
+    in_manifest = in_manifest.map(_add_decoding)
+    logging.info(f"Saving manifest to {out_manifest}")
+    in_manifest.to_file(out_manifest)
+
+
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    cuts = load_manifest_lazy(args.input_manifest)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        splitted_cuts = cuts.split(num_splits=world_size)
+        mp.spawn(
+            run, args=(world_size, args, splitted_cuts), nprocs=world_size, join=True
+        )
+    else:
+        run(rank=0, world_size=1, args=args, cuts=cuts)
+
+
+@torch.no_grad()
+def run(rank, world_size, args, cuts):
+    params = get_params()
+    params.update(vars(args))
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if params.use_pre_text:
+        params.suffix += f"-pre-text-{params.pre_text_transform}"
+
+    if params.use_style_prompt:
+        params.suffix += f"-style-prompt-{params.style_text_transform}"
+
+    params.suffix += f"-{rank}"
+
+    world_size = params.world_size
+
+    params.output_manifest = Path(params.output_manifest)
+    if world_size > 1:
+        cuts = cuts[rank]
+        out_name = params.output_manifest / f"with_decoding_job_{rank}.jsonl.gz"
+    else:
+        out_name = params.output_manifest / "with_decoding.jsonl.gz"
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    setup_logger(f"{params.log_dir}/log-get-manifest-with-decoding-{rank}")
+    logging.info("Decoding started")
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    tokenizer = get_tokenizer(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+
+    LM = None
+
+    decoding_graph = None
+    word_table = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    dl = libriheavy.valid_dataloaders(
+        cuts, text_sampling_func=naive_triplet_text_sampling
+    )
+
+    test_sets = ["test"]
+    test_dl = [dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        biasing_dict = None
+
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            tokenizer=tokenizer,
+            biasing_dict=biasing_dict,
+            word_table=word_table,
+            decoding_graph=decoding_graph,
+        )
+
+        # save_results(
+        #     params=params,
+        #     test_set_name=test_set,
+        #     results_dict=results_dict,
+        # )
+
+        add_decoding_result_to_manifest(
+            in_manifest=cuts,
+            out_manifest=out_name,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+# torch.set_num_threads(1)
+# torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/decoder.py b/egs/libriheavy/ASR/zipformer_prompt_asr/decoder.py
new file mode 100644
index 000000000..93e0f9f7e
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/decoder.py
@@ -0,0 +1,130 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from scaling import Balancer
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        decoder_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          decoder_dim:
+            Dimension of the input embedding, and of the decoder output.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=decoder_dim,
+            padding_idx=blank_id,
+        )
+        # the balancers are to avoid any drift in the magnitude of the
+        # embeddings, which would interact badly with parameter averaging.
+        self.balancer = Balancer(
+            decoder_dim,
+            channel_dim=-1,
+            min_positive=0.0,
+            max_positive=1.0,
+            min_abs=0.5,
+            max_abs=1.0,
+            prob=0.05,
+        )
+
+        self.blank_id = blank_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        self.vocab_size = vocab_size
+
+        if context_size > 1:
+            self.conv = nn.Conv1d(
+                in_channels=decoder_dim,
+                out_channels=decoder_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=decoder_dim // 4,  # group size == 4
+                bias=False,
+            )
+            self.balancer2 = Balancer(
+                decoder_dim,
+                channel_dim=-1,
+                min_positive=0.0,
+                max_positive=1.0,
+                min_abs=0.5,
+                max_abs=1.0,
+                prob=0.05,
+            )
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, decoder_dim).
+        """
+        y = y.to(torch.int64)
+        # this stuff about clamp() is a temporary fix for a mismatch
+        # at utterance start, we use negative ids in beam_search.py
+        embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1)
+
+        embedding_out = self.balancer(embedding_out)
+
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+            embedding_out = F.relu(embedding_out)
+            embedding_out = self.balancer2(embedding_out)
+
+        return embedding_out
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/encoder_interface.py b/egs/libriheavy/ASR/zipformer_prompt_asr/encoder_interface.py
new file mode 100644
index 000000000..257facce4
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/encoder_interface.py
@@ -0,0 +1,43 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+
+
+class EncoderInterface(nn.Module):
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (batch_size, input_seq_len, num_features)
+            containing the input features.
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames
+            in `x` before padding.
+        Returns:
+          Return a tuple containing two tensors:
+            - encoder_out, a tensor of (batch_size, out_seq_len, output_dim)
+              containing unnormalized probabilities, i.e., the output of a
+              linear layer.
+            - encoder_out_lens, a tensor of shape (batch_size,) containing
+              the number of frames in `encoder_out` before padding.
+        """
+        raise NotImplementedError("Please implement it in a subclass")
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/export_PromptASR.py b/egs/libriheavy/ASR/zipformer_prompt_asr/export_PromptASR.py
new file mode 100644
index 000000000..e0bc556a8
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/export_PromptASR.py
@@ -0,0 +1,255 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2023 Xiaomi Corporation (Author: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+
+"""
+Export `model.state_dict()`
+
+- For non-streaming model:
+
+./zipformer_prompt_asr/export_PromptASR.py \
+  --exp-dir ./zipformer_prompt_asr/exp \
+  --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \
+  --epoch 50 \
+  --avg 10
+
+- For streaming model:
+
+./zipformer_prompt_asr/export_PromptASR.py \
+  --exp-dir ./zipformer_prompt_asr/exp \
+  --causal 1 \
+  --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \
+  --epoch 50 \
+  --avg 10
+
+It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
+load it by `icefall.checkpoint.load_checkpoint()`.
+
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+import k2
+import torch
+from torch import Tensor, nn
+from train_bert_encoder import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import make_pad_mask, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        It will generate a file named jit_script.pt.
+        Check ./jit_pretrained.py for how to use it.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    # if torch.cuda.is_available():
+    #     device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.eval()
+
+    assert params.jit is False, "Jit is not supported yet"
+
+    logging.info("Not using torchscript. Export model.state_dict()")
+    # Save it using a format so that it can be loaded
+    # by :func:`load_checkpoint`
+    filename = params.exp_dir / "pretrained.pt"
+    torch.save({"model": model.state_dict()}, str(filename))
+    logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/joiner.py b/egs/libriheavy/ASR/zipformer_prompt_asr/joiner.py
new file mode 100644
index 000000000..59f822748
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/joiner.py
@@ -0,0 +1,86 @@
+# Copyright    2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+from scaling import ScaledLinear
+
+
+class Joiner(nn.Module):
+    def __init__(
+        self,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+        context_dim: int = 512,
+        context_injection: bool = False,
+    ):
+        super().__init__()
+
+        self.encoder_proj = ScaledLinear(encoder_dim, joiner_dim, initial_scale=0.25)
+        self.decoder_proj = ScaledLinear(decoder_dim, joiner_dim, initial_scale=0.25)
+        self.output_linear = nn.Linear(joiner_dim, vocab_size)
+        if context_injection:
+            self.context_proj = ScaledLinear(
+                context_dim, joiner_dim, initial_scale=0.25
+            )
+        else:
+            self.context_proj = None
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+        context: torch.Tensor = None,
+        project_input: bool = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C).
+          context:
+            An embedding vector representing the previous context information
+          project_input:
+            If true, apply input projections encoder_proj and decoder_proj.
+            If this is false, it is the user's responsibility to do this
+            manually.
+        Returns:
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim == 4
+        assert encoder_out.shape[:-1] == decoder_out.shape[:-1]
+
+        if project_input:
+            if context:
+                logit = (
+                    self.encoder_proj(encoder_out)
+                    + self.decoder_proj(decoder_out)
+                    + self.context_proj(context)
+                )
+            else:
+                logit = self.encoder_proj(encoder_out) + self.decoder_proj(decoder_out)
+        else:
+            if context is not None:
+                logit = encoder_out + decoder_out + context.unsqueeze(1).unsqueeze(1)
+            else:
+                logit = encoder_out + decoder_out
+
+        logit = self.output_linear(torch.tanh(logit))
+
+        return logit
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/ls_text_normalization.py b/egs/libriheavy/ASR/zipformer_prompt_asr/ls_text_normalization.py
new file mode 100644
index 000000000..9a693ca4f
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/ls_text_normalization.py
@@ -0,0 +1,153 @@
+import re
+
+words = {
+    0: "zero",
+    1: "one",
+    2: "two",
+    3: "three",
+    4: "four",
+    5: "five",
+    6: "six",
+    7: "seven",
+    8: "eight",
+    9: "nine",
+    10: "ten",
+    11: "eleven",
+    12: "twelve",
+    13: "thirteen",
+    14: "fourteen",
+    15: "fifteen",
+    16: "sixteen",
+    17: "seventeen",
+    18: "eighteen",
+    19: "nineteen",
+    20: "twenty",
+    30: "thirty",
+    40: "forty",
+    50: "fifty",
+    60: "sixty",
+    70: "seventy",
+    80: "eighty",
+    90: "ninety",
+}
+ordinal_nums = [
+    "zeroth",
+    "first",
+    "second",
+    "third",
+    "fourth",
+    "fifth",
+    "sixth",
+    "seventh",
+    "eighth",
+    "ninth",
+    "tenth",
+    "eleventh",
+    "twelfth",
+    "thirteenth",
+    "fourteenth",
+    "fifteenth",
+    "sixteenth",
+    "seventeenth",
+    "eighteenth",
+    "nineteenth",
+    "twentieth",
+]
+
+num_ordinal_dict = {num: ordinal_nums[num] for num in range(21)}
+
+
+def year_to_words(num: int):
+    assert isinstance(num, int), num
+    # check if a num is representing a year
+    if num > 1500 and num < 2000:
+        return words[num // 100] + " " + num_to_words(num % 100)
+    elif num == 2000:
+        return "TWO THOUSAND"
+    elif num > 2000:
+        return "TWO THOUSAND AND " + num_to_words(num % 100)
+    else:
+        return num_to_words(num)
+
+
+def num_to_words(num: int):
+    # Return the English words of a integer number
+
+    # If this is a year number
+    if num > 1500 and num < 2030:
+        return year_to_words(num)
+
+    if num < 20:
+        return words[num]
+    if num < 100:
+        if num % 10 == 0:
+            return words[num // 10 * 10]
+        else:
+            return words[num // 10 * 10] + " " + words[num % 10]
+    if num < 1000:
+        return words[num // 100] + " hundred and " + num_to_words(num % 100)
+    if num < 1000000:
+        return num_to_words(num // 1000) + " thousand " + num_to_words(num % 1000)
+    return num
+
+
+def num_to_ordinal_word(num: int):
+
+    return num_ordinal_dict.get(num, num_to_words(num)).upper()
+
+
+def replace_full_width_symbol(s: str) -> str:
+    # replace full-width symbol with theri half width counterpart
+    s = s.replace("“", '"')
+    s = s.replace("”", '"')
+    s = s.replace("‘", "'")
+    s = s.replace("’", "'")
+
+    return s
+
+
+def decoding_normalization(text: str) -> str:
+    text = replace_full_width_symbol(text)
+
+    # Only keep all alpha-numeric characters, hypen and apostrophe
+    text = text.replace("-", " ")
+    text = re.sub(r"[^a-zA-Z0-9\s']+", "", text)
+    return text
+
+
+def word_normalization(word: str) -> str:
+    # 1 .Use full word for some abbreviation
+    # 2. Convert digits to english words
+    # 3. Convert ordinal number to english words
+    if word == "MRS":
+        return "MISSUS"
+    if word == "MR":
+        return "MISTER"
+    if word == "ST":
+        return "SAINT"
+    if word == "ECT":
+        return "ET CETERA"
+    if word.isnumeric():
+        word = num_to_words(int(word))
+        return str(word).upper()
+    #  e.g 9TH, 6TH
+    if word[-2:] == "TH" and word[0].isnumeric():
+        return num_to_ordinal_word(int(word[:-2])).upper()
+    if word[0] == "'":
+        return word[1:]
+
+    return word
+
+
+def simple_normalization(text: str) -> str:
+    text = replace_full_width_symbol(text)
+    text = text.replace("--", " ")
+
+    return text
+
+
+if __name__ == "__main__":
+
+    s = str(1830)
+    out = word_normalization(s)
+    print(s, out)
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/model_baseline.py b/egs/libriheavy/ASR/zipformer_prompt_asr/model_baseline.py
new file mode 100644
index 000000000..77b4057c4
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/model_baseline.py
@@ -0,0 +1,262 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import random
+import warnings
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear, penalize_abs_values_gt
+from torch import Tensor
+
+from icefall.utils import add_sos, make_pad_mask
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder_embed: nn.Module,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+    ):
+        """
+        Args:
+          encoder_embed:
+            It is a Convolutional 2D subsampling module. It converts
+            an input of shape (N, T, idim) to an output of of shape
+            (N, T', odim), where T' = (T-3)//2-2 = (T-7)//2.
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder_embed = encoder_embed
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = ScaledLinear(
+            encoder_dim,
+            vocab_size,
+            initial_scale=0.25,
+        )
+        self.simple_lm_proj = ScaledLinear(
+            decoder_dim,
+            vocab_size,
+            initial_scale=0.25,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          text:
+            A 2-D tensor of integer dtype containing prompt text, of shape (N, T).
+            It is exptected to contain the style prompt (first) and then the content
+            prompt.
+          text_lens:
+            A 1-D tensor of shape (N,). It contains the number of elements (bytes)
+            in `text` before padding, which will include the lengths of the
+            style plus the content prompt.
+          style_lens:
+            A 1-D tensor of shape (N,), containing the number of elements (bytes)
+            within each row of `text` that correspond to the style prompt (these
+            are expected to come first).
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        x, x_lens = self.encoder_embed(x, x_lens)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, x_lens = self.encoder(
+            x,
+            x_lens,
+            src_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros(
+            (encoder_out.size(0), 4),
+            dtype=torch.int64,
+            device=encoder_out.device,
+        )
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = self.joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss)
+
+    def encode_audio(
+        self,
+        feature: Tensor,
+        feature_lens: Tensor,
+        memory: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        """Encode the input audio features
+
+        Args:
+            feature (Tensor): Input audio (N,T,C)
+            feature_lens (Tensor): Length of input audio (N,)
+        Returns:
+            Tuple[Tensor, Tensor]: Encoded acoustic features and length
+        """
+        x, x_lens = self.encoder_embed(feature, feature_lens)
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(
+            x=x,
+            x_lens=x_lens,
+            src_key_padding_mask=src_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return encoder_out, encoder_out_lens
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/model_with_BERT.py b/egs/libriheavy/ASR/zipformer_prompt_asr/model_with_BERT.py
new file mode 100644
index 000000000..21c7b4fac
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/model_with_BERT.py
@@ -0,0 +1,392 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import random
+import warnings
+from typing import Dict, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear, penalize_abs_values_gt
+from torch import Tensor
+
+from icefall.utils import add_sos, make_pad_mask
+
+
+class PromptedTransducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder_embed: nn.Module,
+        encoder: EncoderInterface,
+        text_encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+        use_BERT: bool = True,
+        text_encoder_type: str = "BERT",
+        text_encoder_adapter: bool = False,
+        freeze_text_encoder: bool = True,
+        context_fuser: nn.Module = None,
+    ):
+        """
+        Args:
+          encoder_embed:
+            It is a Convolutional 2D subsampling module. It converts
+            an input of shape (N, T, idim) to an output of of shape
+            (N, T', odim), where T' = (T-3)//2-2 = (T-7)//2.
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          text_encoder:
+            This is a encoder that processes text information (e.g content prompt
+            and style prompt). The input is `x` of (N,T) and `x_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
+            Its output shape is (N, T, U, vocab_size). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+          text_encoder_type:
+            The type of the text_encoder. Supported are (BERT, DistilBERT)
+          context_fuser
+            A optional module that fuses the embeddings of text encoder. The fused embedding
+            will be added to the joiner.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder_embed = encoder_embed
+        self.encoder = encoder
+        self.text_encoder = text_encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = ScaledLinear(
+            encoder_dim,
+            vocab_size,
+            initial_scale=0.25,
+        )
+        self.simple_lm_proj = ScaledLinear(
+            decoder_dim,
+            vocab_size,
+            initial_scale=0.25,
+        )
+
+        self.use_BERT = use_BERT  # if the text encoder is a pre-trained BERT
+        self.context_fuser = context_fuser
+
+        assert text_encoder_type in (
+            "BERT",
+            "DistilBERT",
+            "BERT-UNCASED",
+        ), f"Unseen text_encoder type {text_encoder_type}"
+        self.text_encoder_dim = (
+            self.text_encoder.config.hidden_size
+            if text_encoder_type in ("BERT", "BERT-UNCASED")
+            else self.text_encoder.config.dim
+        )
+        self.freeze_text_encoder = freeze_text_encoder
+
+        if text_encoder_adapter:
+            self.text_encoder_adapter = nn.Sequential(
+                nn.Linear(self.text_encoder_dim, self.text_encoder_dim, bias=False),
+                nn.Tanh(),
+            )
+        else:
+            self.text_encoder_adapter = None
+
+        self.style_prompt_embedding = nn.Parameter(
+            torch.full((self.text_encoder_dim,), 0.5)
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        encoded_inputs: Dict,
+        style_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        use_pre_text: bool = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          text:
+            A 2-D tensor of integer dtype containing prompt text, of shape (N, T).
+            It is exptected to contain the style prompt (first) and then the content
+            prompt.
+          text_lens:
+            A 1-D tensor of shape (N,). It contains the number of elements (bytes)
+            in `text` before padding, which will include the lengths of the
+            style plus the content prompt.
+          style_lens:
+            A 1-D tensor of shape (N,), containing the number of elements (bytes)
+            within each row of `text` that correspond to the style prompt (these
+            are expected to come first).
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        if self.freeze_text_encoder:
+            self.text_encoder.eval()
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        x, x_lens = self.encoder_embed(x, x_lens)
+
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # freeze the BERT text encoder
+
+        if use_pre_text:
+            memory, memory_key_padding_mask = self.encode_text(
+                encoded_inputs, style_lens=style_lens
+            )
+        else:
+            memory = None
+            memory_key_padding_mask = None
+
+        encoder_out, x_lens = self.encoder(
+            x,
+            x_lens,
+            src_key_padding_mask,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros(
+            (encoder_out.size(0), 4),
+            dtype=torch.int64,
+            device=encoder_out.device,
+        )
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=self.joiner.encoder_proj(encoder_out),
+            lm=self.joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        if self.context_fuser is not None and memory is not None:
+            memory = memory.permute(1, 0, 2)  # (T,N,C) -> (N,T,C)
+            context = self.context_fuser(memory, padding_mask=memory_key_padding_mask)
+            context = self.joiner.context_proj(context)
+        else:
+            context = None
+
+        logits = self.joiner(am_pruned, lm_pruned, context=context, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss)
+
+    def _add_style_indicator(self, memory: Tensor, style_lens: Tensor):
+        """
+        Adds to `memory` an indicator that is 1.0 for positions that correspond to
+        the `style prompt` and 0 elsewhere.  The scale can be fixed because the
+        scale of the embedding vector can adjust to compensate.
+
+        Args:
+            memory: (memory_len, batch_size, embed_dim)
+            style_lens: (batch_size,),  a vector of lengths of the style prompt.
+        """
+
+        (memory_len, batch_size, embed_dim) = memory.shape
+
+        indicator = (
+            torch.arange(memory_len, device=memory.device).unsqueeze(-1) < style_lens
+        )
+        indicator = indicator.to(memory.dtype)
+
+        extra_term = torch.zeros_like(memory)
+        extra_term += indicator.unsqueeze(-1) * self.style_prompt_embedding.expand(
+            memory_len, batch_size, self.text_encoder_dim
+        )
+
+        return memory + extra_term
+
+    def encode_text(
+        self,
+        encoded_inputs: Dict,
+        style_lens: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """Get the embeddings of text
+
+        Args:
+            encoded_inputs: The encoded inputs generated by a tokenizer (Dict)
+
+        Returns:
+            Tuple[Tensor, Tensor]: Returns the text embeddings encoded by the
+            text_encoder and the attention mask
+        """
+        text_lens = encoded_inputs.pop("length")  # need to use pop to remove this item
+
+        # Freeze the pre-trained text encoder
+        with torch.no_grad():
+            memory = self.text_encoder(**encoded_inputs)["last_hidden_state"]  # (B,T,C)
+            memory = memory.permute(1, 0, 2)
+
+        # Text encoder adapter
+        if self.text_encoder_adapter is not None:
+            memory = self.text_encoder_adapter(memory)
+
+        memory = self._add_style_indicator(memory, style_lens)
+
+        memory_key_padding_mask = make_pad_mask(text_lens)
+
+        return memory, memory_key_padding_mask
+
+    def encode_audio(
+        self,
+        feature: Tensor,
+        feature_lens: Tensor,
+        memory: Optional[Tensor],
+        memory_key_padding_mask: Optional[Tensor],
+    ) -> Tuple[Tensor, Tensor]:
+        """Encode the input audio features
+
+        Args:
+            feature (Tensor): Input audio (N,T,C)
+            feature_lens (Tensor): Length of input audio (N,)
+            memory (Tensor): Embeddings from the text encoder
+            memory_key_padding_mask (Tensor): _description_
+
+        Returns:
+            Tuple[Tensor, Tensor]: _description_
+        """
+        x, x_lens = self.encoder_embed(feature, feature_lens)
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        encoder_out, encoder_out_lens = self.encoder(
+            x=x,
+            x_lens=x_lens,
+            src_key_padding_mask=src_key_padding_mask,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )
+        encoder_out = encoder_out.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return encoder_out, encoder_out_lens
+
+
+Transducer = PromptedTransducer  # for decoding
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/optim.py b/egs/libriheavy/ASR/zipformer_prompt_asr/optim.py
new file mode 100644
index 000000000..a767761eb
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/optim.py
@@ -0,0 +1,1168 @@
+#      Copyright      2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import contextlib
+import logging
+import random
+from collections import defaultdict
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+from lhotse.utils import fix_random_seed
+from torch import Tensor
+from torch.optim import Optimizer
+
+
+class BatchedOptimizer(Optimizer):
+    """
+    This class adds to class Optimizer the capability to optimize parameters in batches:
+    it will stack the parameters and their grads for you so the optimizer can work
+    on tensors with an extra leading dimension.  This is intended for speed with GPUs,
+    as it reduces the number of kernels launched in the optimizer.
+
+    Args:
+      params:
+    """
+
+    def __init__(self, params, defaults):
+        super(BatchedOptimizer, self).__init__(params, defaults)
+
+    @contextlib.contextmanager
+    def batched_params(self, param_group, group_params_names):
+        """
+        This function returns (technically, yields) a list of
+          of tuples (p, state), where
+        p is a `fake` parameter that is stacked (over axis 0) from real parameters
+        that share the same shape, and its gradient is also stacked;
+        `state` is the state corresponding to this batch of parameters
+        (it will be physically located in the "state" for one of the real
+        parameters, the last one that has any particular shape and dtype).
+
+        This function is decorated as a context manager so that it can
+        write parameters back to their "real" locations.
+
+        The idea is, instead of doing:
+        <code>
+          for p in group["params"]:
+             state = self.state[p]
+             ...
+        </code>
+        you can do:
+        <code>
+          with self.batched_params(group["params"]) as batches:
+             for p, state, p_names in batches:
+                 ...
+        </code>
+
+        Args:
+          group: a parameter group, which is a list of parameters; should be
+                one of self.param_groups.
+          group_params_names: name for each parameter in group,
+                which is List[str].
+        """
+        batches = defaultdict(
+            list
+        )  # `batches` maps from tuple (dtype_as_str,*shape) to list of nn.Parameter
+        batches_names = defaultdict(
+            list
+        )  # `batches` maps from tuple (dtype_as_str,*shape) to list of str
+
+        assert len(param_group) == len(group_params_names)
+        for p, named_p in zip(param_group, group_params_names):
+            key = (str(p.dtype), *p.shape)
+            batches[key].append(p)
+            batches_names[key].append(named_p)
+
+        batches_names_keys = list(batches_names.keys())
+        sorted_idx = sorted(
+            range(len(batches_names)), key=lambda i: batches_names_keys[i]
+        )
+        batches_names = [batches_names[batches_names_keys[idx]] for idx in sorted_idx]
+        batches = [batches[batches_names_keys[idx]] for idx in sorted_idx]
+
+        stacked_params_dict = dict()
+
+        # turn batches into a list, in deterministic order.
+        # tuples will contain tuples of (stacked_param, state, stacked_params_names),
+        # one for each batch in `batches`.
+        tuples = []
+
+        for batch, batch_names in zip(batches, batches_names):
+            p = batch[0]
+            # we arbitrarily store the state in the
+            # state corresponding to the 1st parameter in the
+            # group.  class Optimizer will take care of saving/loading state.
+            state = self.state[p]
+            p_stacked = torch.stack(batch)
+            grad = torch.stack(
+                [torch.zeros_like(p) if p.grad is None else p.grad for p in batch]
+            )
+            p_stacked.grad = grad
+            stacked_params_dict[key] = p_stacked
+            tuples.append((p_stacked, state, batch_names))
+
+        yield tuples  # <-- calling code will do the actual optimization here!
+
+        for ((stacked_params, _state, _names), batch) in zip(tuples, batches):
+            for i, p in enumerate(batch):  # batch is list of Parameter
+                p.copy_(stacked_params[i])
+
+
+class ScaledAdam(BatchedOptimizer):
+    """
+     Implements 'Scaled Adam', a variant of Adam where we scale each parameter's update
+     proportional to the norm of that parameter; and also learn the scale of the parameter,
+     in log space, subject to upper and lower limits (as if we had factored each parameter as
+     param = underlying_param * log_scale.exp())
+
+
+     Args:
+          params:  The parameters or param_groups to optimize (like other Optimizer subclasses)
+                   Unlike common optimizers, which accept model.parameters() or groups of parameters(),
+                   this optimizer could accept model.named_parameters() or groups of named_parameters().
+                   See comments of function _get_names_of_parameters for its 4 possible cases.
+              lr:  The learning rate.  We will typically use a learning rate schedule that starts
+                   at 0.03 and decreases over time, i.e. much higher than other common
+                   optimizers.
+     clipping_scale: (e.g. 2.0)
+                   A scale for gradient-clipping: if specified, the normalized gradients
+                   over the whole model will be clipped to have 2-norm equal to
+                   `clipping_scale` times the median 2-norm over the most recent period
+                   of `clipping_update_period` minibatches.  By "normalized gradients",
+                   we mean after multiplying by the rms parameter value for this tensor
+                   [for non-scalars]; this is appropriate because our update is scaled
+                   by this quantity.
+            betas: beta1,beta2 are momentum constants for regular momentum, and moving sum-sq grad.
+                   Must satisfy 0 < beta <= beta2 < 1.
+     scalar_lr_scale: A scaling factor on the learning rate, that we use to update the
+                   scale of each parameter tensor and scalar parameters of the mode..
+                   If each parameter were decomposed
+                   as p * p_scale.exp(), where (p**2).mean().sqrt() == 1.0, scalar_lr_scale
+                   would be a the scaling factor on the learning rate of p_scale.
+              eps:  A general-purpose epsilon to prevent division by zero
+    param_min_rms: Minimum root-mean-square value of parameter tensor, for purposes of
+                   learning the scale on the parameters (we'll constrain the rms of each non-scalar
+                   parameter tensor to be >= this value)
+    param_max_rms: Maximum root-mean-square value of parameter tensor, for purposes of
+                   learning the scale on the parameters (we'll constrain the rms of each non-scalar
+                   parameter tensor to be <= this value)
+       scalar_max: Maximum absolute value for scalar parameters (applicable if your
+                   model has any parameters with numel() == 1).
+    size_update_period: The periodicity, in steps, with which we update the size (scale)
+                   of the parameter tensor.  This is provided to save a little time
+                   in the update.
+     clipping_update_period: if clipping_scale is specified, this is the period
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=3e-02,
+        clipping_scale=None,
+        betas=(0.9, 0.98),
+        scalar_lr_scale=0.1,
+        eps=1.0e-08,
+        param_min_rms=1.0e-05,
+        param_max_rms=3.0,
+        scalar_max=10.0,
+        size_update_period=4,
+        clipping_update_period=100,
+    ):
+
+        defaults = dict(
+            lr=lr,
+            clipping_scale=clipping_scale,
+            betas=betas,
+            scalar_lr_scale=scalar_lr_scale,
+            eps=eps,
+            param_min_rms=param_min_rms,
+            param_max_rms=param_max_rms,
+            scalar_max=scalar_max,
+            size_update_period=size_update_period,
+            clipping_update_period=clipping_update_period,
+        )
+
+        # If params only contains parameters or group of parameters,
+        # i.e when parameter names are not given,
+        # this flag will be set to False in funciton _get_names_of_parameters.
+        self.show_dominant_parameters = True
+        param_groups, parameters_names = self._get_names_of_parameters(params)
+        super(ScaledAdam, self).__init__(param_groups, defaults)
+        assert len(self.param_groups) == len(parameters_names)
+        self.parameters_names = parameters_names
+
+    def _get_names_of_parameters(
+        self, params_or_named_params
+    ) -> Tuple[List[Dict], List[List[str]]]:
+        """
+        Args:
+          params_or_named_params: according to the way ScaledAdam is initialized in train.py,
+            this argument could be one of following 4 cases,
+            case 1, a generator of parameter, e.g.:
+              optimizer = ScaledAdam(model.parameters(), lr=params.base_lr, clipping_scale=3.0)
+
+            case 2, a list of parameter groups with different config, e.g.:
+              model_param_groups = [
+                      {'params': model.encoder.parameters(), 'lr': 0.05},
+                      {'params': model.decoder.parameters(), 'lr': 0.01},
+                      {'params': model.joiner.parameters(), 'lr': 0.03},
+                      ]
+              optimizer = ScaledAdam(model_param_groups, lr=params.base_lr, clipping_scale=3.0)
+
+            case 3, a generator of named_parameter, e.g.:
+              optimizer = ScaledAdam(model.named_parameters(), lr=params.base_lr, clipping_scale=3.0)
+
+            case 4, a list of named_parameter groups with different config, e.g.:
+              model_named_param_groups = [
+                      {'named_params': model.encoder.named_parameters(), 'lr': 0.05},
+                      {'named_params': model.decoder.named_parameters(), 'lr': 0.01},
+                      {'named_params': model.joiner.named_parameters(), 'lr': 0.03},
+                      ]
+              optimizer = ScaledAdam(model_named_param_groups, lr=params.base_lr, clipping_scale=3.0)
+
+          For case 1 and case 2, input params is used to initialize the underlying torch.optimizer.
+          For case 3 and case 4, firstly, names and params are extracted from input named_params,
+            then, these extracted params are used to initialize the underlying torch.optimizer,
+            and these extracted names are mainly used by function
+            `_show_gradient_dominating_parameter`
+
+        Returns:
+          Returns a tuple containing 2 elements:
+            - `param_groups` with type List[Dict], each Dict element is a parameter group.
+              An example of `param_groups` could be:
+              [
+                  {'params': `one iterable of Parameter`, 'lr': 0.05},
+                  {'params': `another iterable of Parameter`, 'lr': 0.08},
+                  {'params': `a third iterable of Parameter`, 'lr': 0.1},
+              ]
+            - `param_gruops_names` with type List[List[str]],
+               each `List[str]` is for a group['params'] in param_groups,
+               and each `str` is the name of a parameter.
+               A dummy name "foo" is related to each parameter,
+               if input are params without names, i.e. case 1 or case 2.
+        """
+        # variable naming convention in this function:
+        #   p is short for param.
+        #   np is short for named_param.
+        #   p_or_np is short for param_or_named_param.
+        #   cur is short for current.
+        #   group is a dict, e.g. {'params': iterable of parameter, 'lr': 0.05, other fields}.
+        #   groups is a List[group]
+
+        iterable_or_groups = list(params_or_named_params)
+        if len(iterable_or_groups) == 0:
+            raise ValueError("optimizer got an empty parameter list")
+
+        # The first value of returned tuple.  A list of dicts containing at
+        # least 'params' as a key.
+        param_groups = []
+
+        # The second value of returned tuple,
+        # a List[List[str]], each sub-List is for a group.
+        param_groups_names = []
+
+        if not isinstance(iterable_or_groups[0], dict):
+            # case 1 or case 3,
+            # the input is an iterable of parameter or named parameter.
+            param_iterable_cur_group = []
+            param_names_cur_group = []
+            for p_or_np in iterable_or_groups:
+                if isinstance(p_or_np, tuple):
+                    # case 3
+                    name, param = p_or_np
+                else:
+                    # case 1
+                    assert isinstance(p_or_np, torch.Tensor)
+                    param = p_or_np
+                    # Assign a dummy name as a placeholder
+                    name = "foo"
+                    self.show_dominant_parameters = False
+                param_iterable_cur_group.append(param)
+                param_names_cur_group.append(name)
+            param_groups.append({"params": param_iterable_cur_group})
+            param_groups_names.append(param_names_cur_group)
+        else:
+            # case 2 or case 4
+            # the input is groups of parameter or named parameter.
+            for cur_group in iterable_or_groups:
+                assert "named_params" in cur_group
+                name_list = [x[0] for x in cur_group["named_params"]]
+                p_list = [x[1] for x in cur_group["named_params"]]
+                del cur_group["named_params"]
+                cur_group["params"] = p_list
+                param_groups.append(cur_group)
+                param_groups_names.append(name_list)
+
+        return param_groups, param_groups_names
+
+    def __setstate__(self, state):
+        super(ScaledAdam, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        batch = True
+
+        for group, group_params_names in zip(self.param_groups, self.parameters_names):
+
+            with self.batched_params(group["params"], group_params_names) as batches:
+
+                # batches is list of pairs (stacked_param, state).  stacked_param is like
+                # a regular parameter, and will have a .grad, but the 1st dim corresponds to
+                # a stacking dim, it is not a real dim.
+
+                if (
+                    len(batches[0][1]) == 0
+                ):  # if len(first state) == 0: not yet initialized
+                    clipping_scale = 1
+                else:
+                    clipping_scale = self._get_clipping_scale(group, batches)
+
+                for p, state, _ in batches:
+                    # Perform optimization step.
+                    # grad is not going to be None, we handled that when creating the batches.
+                    grad = p.grad
+                    if grad.is_sparse:
+                        raise RuntimeError(
+                            "ScaledAdam optimizer does not support sparse gradients"
+                        )
+                    # State initialization
+                    if len(state) == 0:
+                        self._init_state(group, p, state)
+
+                    self._step_one_batch(group, p, state, clipping_scale)
+
+        return loss
+
+    def _init_state(self, group: dict, p: Tensor, state: dict):
+        """
+        Initializes state dict for parameter 'p'.  Assumes that dim 0 of tensor p
+        is actually the batch dimension, corresponding to batched-together
+        parameters of a given shape.
+
+
+        Args:
+           group:   Dict to look up configuration values.
+               p: The parameter that we are initializing the state for
+           state: Dict from string to whatever state we are initializing
+        """
+        size_update_period = group["size_update_period"]
+
+        state["step"] = 0
+
+        kwargs = {"device": p.device, "dtype": p.dtype}
+
+        # 'delta' implements conventional momentum.  There are
+        # several different kinds of update going on, so rather than
+        # compute "exp_avg" like in Adam, we store and decay a
+        # parameter-change "delta", which combines all forms of
+        # update.  this is equivalent to how it's done in Adam,
+        # except for the first few steps.
+        state["delta"] = torch.zeros_like(p, memory_format=torch.preserve_format)
+
+        batch_size = p.shape[0]
+        numel = p.numel() // batch_size
+        numel = p.numel()
+
+        if numel > 1:
+            # "param_rms" just periodically records the scalar root-mean-square value of
+            # the parameter tensor.
+            # it has a shape like (batch_size, 1, 1, 1, 1)
+            param_rms = (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt()
+            state["param_rms"] = param_rms
+
+            state["scale_exp_avg_sq"] = torch.zeros_like(param_rms)
+            state["scale_grads"] = torch.zeros(
+                size_update_period, *param_rms.shape, **kwargs
+            )
+
+        # exp_avg_sq is the weighted sum of scaled gradients. as in Adam.
+        state["exp_avg_sq"] = torch.zeros_like(p, memory_format=torch.preserve_format)
+
+    def _get_clipping_scale(
+        self, group: dict, tuples: List[Tuple[Tensor, dict, List[str]]]
+    ) -> float:
+        """
+        Returns a scalar factor <= 1.0 that dictates gradient clipping, i.e. we will scale the gradients
+        by this amount before applying the rest of the update.
+
+        Args:
+           group: the parameter group, an item in self.param_groups
+           tuples: a list of tuples of (param, state, param_names)
+                where param is a batched set of parameters,
+                with a .grad (1st dim is batch dim)
+                and state is the state-dict where optimization parameters are kept.
+                param_names is a List[str] while each str is name for a parameter
+                in batched set of parameters "param".
+        """
+        assert len(tuples) >= 1
+        clipping_scale = group["clipping_scale"]
+        (first_p, first_state, _) = tuples[0]
+        step = first_state["step"]
+        if clipping_scale is None or step == 0:
+            # no clipping.  return early on step == 0 because the other
+            # parameters' state won't have been initialized yet.
+            return 1.0
+        clipping_update_period = group["clipping_update_period"]
+
+        tot_sumsq = torch.tensor(0.0, device=first_p.device)
+        for (p, state, param_names) in tuples:
+            grad = p.grad
+            if grad.is_sparse:
+                raise RuntimeError(
+                    "ScaledAdam optimizer does not support sparse gradients"
+                )
+            if p.numel() == p.shape[0]:  # a batch of scalars
+                tot_sumsq += (grad**2).sum()  # sum() to change shape [1] to []
+            else:
+                tot_sumsq += ((grad * state["param_rms"]) ** 2).sum()
+
+        tot_norm = tot_sumsq.sqrt()
+        if "model_norms" not in first_state:
+            first_state["model_norms"] = torch.zeros(
+                clipping_update_period, device=p.device
+            )
+        first_state["model_norms"][step % clipping_update_period] = tot_norm
+
+        if step % clipping_update_period == 0:
+            # Print some stats.
+            # We don't reach here if step == 0 because we would have returned
+            # above.
+            sorted_norms = first_state["model_norms"].sort()[0].to("cpu")
+            quartiles = []
+            for n in range(0, 5):
+                index = min(
+                    clipping_update_period - 1, (clipping_update_period // 4) * n
+                )
+                quartiles.append(sorted_norms[index].item())
+
+            median = quartiles[2]
+            threshold = clipping_scale * median
+            first_state["model_norm_threshold"] = threshold
+            percent_clipped = (
+                first_state["num_clipped"] * 100.0 / clipping_update_period
+                if "num_clipped" in first_state
+                else 0.0
+            )
+            first_state["num_clipped"] = 0
+            quartiles = " ".join(["%.3e" % x for x in quartiles])
+            logging.info(
+                f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, "
+                f"threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}"
+            )
+
+        if step < clipping_update_period:
+            return 1.0  # We have not yet estimated a norm to clip to.
+        else:
+            try:
+                model_norm_threshold = first_state["model_norm_threshold"]
+            except KeyError:
+                logging.info(
+                    "Warning: model_norm_threshold not in state: possibly "
+                    "you changed config when restarting, adding clipping_scale option?"
+                )
+                return 1.0
+            ans = min(1.0, (model_norm_threshold / (tot_norm + 1.0e-20)).item())
+            if ans < 1.0:
+                first_state["num_clipped"] += 1
+            if ans < 0.1:
+                logging.warn(
+                    f"Scaling gradients by {ans}, model_norm_threshold={model_norm_threshold}"
+                )
+                if self.show_dominant_parameters:
+                    assert p.shape[0] == len(param_names)
+                    self._show_gradient_dominating_parameter(tuples, tot_sumsq)
+            return ans
+
+    def _show_gradient_dominating_parameter(
+        self, tuples: List[Tuple[Tensor, dict, List[str]]], tot_sumsq: Tensor
+    ):
+        """
+        Show information of parameter which dominates tot_sumsq.
+
+        Args:
+           tuples: a list of tuples of (param, state, param_names)
+                where param is a batched set of parameters,
+                with a .grad (1st dim is batch dim)
+                and state is the state-dict where optimization parameters are kept.
+                param_names is a List[str] while each str is name for a parameter
+                in batched set of parameters "param".
+            tot_sumsq: sumsq of all parameters. Though it's could be calculated
+                from tuples, we still pass it to save some time.
+        """
+        all_sumsq_orig = {}
+        for (p, state, batch_param_names) in tuples:
+            # p is a stacked batch parameters.
+            batch_grad = p.grad
+            if p.numel() == p.shape[0]:  # a batch of scalars
+                batch_sumsq_orig = batch_grad**2
+                # Dummy values used by following `zip` statement.
+                batch_rms_orig = torch.ones(p.shape[0])
+            else:
+                batch_rms_orig = state["param_rms"]
+                batch_sumsq_orig = ((batch_grad * batch_rms_orig) ** 2).sum(
+                    dim=list(range(1, batch_grad.ndim))
+                )
+
+            for name, sumsq_orig, rms, grad in zip(
+                batch_param_names, batch_sumsq_orig, batch_rms_orig, batch_grad
+            ):
+
+                proportion_orig = sumsq_orig / tot_sumsq
+                all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad)
+
+        assert torch.isclose(
+            sum([value[0] for value in all_sumsq_orig.values()]).cpu(),
+            torch.tensor(1.0),
+        )
+        sorted_by_proportion = {
+            k: v
+            for k, v in sorted(
+                all_sumsq_orig.items(), key=lambda item: item[1][0], reverse=True
+            )
+        }
+        dominant_param_name = next(iter(sorted_by_proportion))
+        (
+            dominant_proportion,
+            dominant_sumsq,
+            dominant_rms,
+            dominant_grad,
+        ) = sorted_by_proportion[dominant_param_name]
+        logging.info(
+            f"Parameter dominating tot_sumsq {dominant_param_name}"
+            f" with proportion {dominant_proportion:.2f},"
+            f" where dominant_sumsq=(grad_sumsq*orig_rms_sq)"
+            f"={dominant_sumsq:.3e},"
+            f" grad_sumsq={(dominant_grad**2).sum():.3e},"
+            f" orig_rms_sq={(dominant_rms**2).item():.3e}"
+        )
+
+    def _step_one_batch(
+        self, group: dict, p: Tensor, state: dict, clipping_scale: float
+    ):
+        """
+        Do the step for one parameter, which is actually going to be a batch of
+        `real` parameters, with dim 0 as the batch dim.
+        Args:
+                  group:  dict to look up configuration values
+                    p: parameter to update (actually multiple parameters stacked together
+                       as a batch)
+                  state: state-dict for p, to look up the optimizer state
+        """
+        lr = group["lr"]
+        size_update_period = group["size_update_period"]
+        beta1 = group["betas"][0]
+
+        grad = p.grad
+        if clipping_scale != 1.0:
+            grad = grad * clipping_scale
+        step = state["step"]
+        delta = state["delta"]
+
+        delta.mul_(beta1)
+        batch_size = p.shape[0]
+        numel = p.numel() // batch_size
+        if numel > 1:
+            # Update the size/scale of p, and set param_rms
+            scale_grads = state["scale_grads"]
+            scale_grads[step % size_update_period] = (p * grad).sum(
+                dim=list(range(1, p.ndim)), keepdim=True
+            )
+            if step % size_update_period == size_update_period - 1:
+                param_rms = state["param_rms"]  # shape: (batch_size, 1, 1, ..)
+                param_rms.copy_(
+                    (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt()
+                )
+                if step > 0:
+                    # self._size_update() learns the overall scale on the
+                    # parameter, by shrinking or expanding it.
+                    self._size_update(group, scale_grads, p, state)
+
+        if numel == 1:
+            # For parameters with 1 element we just use regular Adam.
+            # Updates delta.
+            self._step_scalar(group, p, state)
+        else:
+            self._step(group, p, state)
+
+        state["step"] = step + 1
+
+    def _size_update(
+        self, group: dict, scale_grads: Tensor, p: Tensor, state: dict
+    ) -> None:
+        """
+               Called only where p.numel() > 1, this updates the scale of the parameter.
+               If we imagine: p =  underlying_param * scale.exp(), and we are doing
+               gradient descent on underlying param and on scale, this function does the update
+               on `scale`.
+
+               Args:
+              group: dict to look up configuration values
+        scale_grads: a tensor of shape (size_update_period, batch_size, 1, 1,...) containing
+                      grads w.r.t. the scales.
+                  p:  The parameter to update
+               state: The state-dict of p
+        """
+
+        param_rms = state["param_rms"]
+        beta1, beta2 = group["betas"]
+        size_lr = group["lr"] * group["scalar_lr_scale"]
+        param_min_rms = group["param_min_rms"]
+        param_max_rms = group["param_max_rms"]
+        eps = group["eps"]
+        step = state["step"]
+        batch_size = p.shape[0]
+
+        size_update_period = scale_grads.shape[0]
+        # correct beta2 for the size update period: we will have
+        # faster decay at this level.
+        beta2_corr = beta2**size_update_period
+
+        scale_exp_avg_sq = state["scale_exp_avg_sq"]  # shape: (batch_size, 1, 1, ..)
+        scale_exp_avg_sq.mul_(beta2_corr).add_(
+            (scale_grads**2).mean(dim=0),  # mean over dim `size_update_period`
+            alpha=1 - beta2_corr,
+        )  # shape is (batch_size, 1, 1, ...)
+
+        # The 1st time we reach here is when size_step == 1.
+        size_step = (step + 1) // size_update_period
+        bias_correction2 = 1 - beta2_corr**size_step
+        # we don't bother with bias_correction1; this will help prevent divergence
+        # at the start of training.
+
+        denom = scale_exp_avg_sq.sqrt() + eps
+
+        scale_step = (
+            -size_lr * (bias_correction2**0.5) * scale_grads.sum(dim=0) / denom
+        )
+
+        is_too_small = param_rms < param_min_rms
+
+        # when the param gets too small, just don't shrink it any further.
+        scale_step.masked_fill_(is_too_small, 0.0)
+
+        # and ensure the parameter rms after update never exceeds param_max_rms.
+        # We have to look at the trained model for parameters at or around the
+        # param_max_rms, because sometimes they can indicate a problem with the
+        # topology or settings.
+        scale_step = torch.minimum(scale_step, (param_max_rms - param_rms) / param_rms)
+
+        delta = state["delta"]
+        # the factor of (1-beta1) relates to momentum.
+        delta.add_(p * scale_step, alpha=(1 - beta1))
+
+    def _step(self, group: dict, p: Tensor, state: dict):
+        """
+        This function does the core update of self.step(), in the case where the members of
+        the batch have more than 1 element.
+
+        Args:
+            group: A dict which will be used to look up configuration values
+                p: The parameter to be updated
+             grad: The grad of p
+            state: The state-dict corresponding to parameter p
+
+        This function modifies p.
+        """
+        grad = p.grad
+        lr = group["lr"]
+        beta1, beta2 = group["betas"]
+        eps = group["eps"]
+        param_min_rms = group["param_min_rms"]
+        step = state["step"]
+
+        exp_avg_sq = state["exp_avg_sq"]
+        exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=(1 - beta2))
+
+        this_step = state["step"] - (state["zero_step"] if "zero_step" in state else 0)
+        bias_correction2 = 1 - beta2 ** (this_step + 1)
+        if bias_correction2 < 0.99:
+            # note: not in-place.
+            exp_avg_sq = exp_avg_sq * (1.0 / bias_correction2)
+
+        denom = exp_avg_sq.sqrt()
+        denom += eps
+        grad = grad / denom
+
+        alpha = -lr * (1 - beta1) * state["param_rms"].clamp(min=param_min_rms)
+
+        delta = state["delta"]
+        delta.add_(grad * alpha)
+        p.add_(delta)
+
+    def _step_scalar(self, group: dict, p: Tensor, state: dict):
+        """
+        A simplified form of the core update for scalar tensors, where we cannot get a good
+        estimate of the parameter rms.
+        """
+        beta1, beta2 = group["betas"]
+        scalar_max = group["scalar_max"]
+        eps = group["eps"]
+        lr = group["lr"] * group["scalar_lr_scale"]
+        grad = p.grad
+
+        exp_avg_sq = state["exp_avg_sq"]  # shape: (batch_size,)
+        exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+
+        # bias_correction2 is like in Adam.  Don't bother with bias_correction1;
+        # slower update at the start will help stability anyway.
+        bias_correction2 = 1 - beta2 ** (state["step"] + 1)
+        denom = (exp_avg_sq / bias_correction2).sqrt() + eps
+
+        delta = state["delta"]
+        delta.add_(grad / denom, alpha=-lr * (1 - beta1))
+        p.clamp_(min=-scalar_max, max=scalar_max)
+        p.add_(delta)
+
+
+class LRScheduler(object):
+    """
+    Base-class for learning rate schedulers where the learning-rate depends on both the
+    batch and the epoch.
+    """
+
+    def __init__(self, optimizer: Optimizer, verbose: bool = False):
+        # Attach optimizer
+        if not isinstance(optimizer, Optimizer):
+            raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__))
+        self.optimizer = optimizer
+        self.verbose = verbose
+
+        for group in optimizer.param_groups:
+            group.setdefault("base_lr", group["lr"])
+
+        self.base_lrs = [group["base_lr"] for group in optimizer.param_groups]
+
+        self.epoch = 0
+        self.batch = 0
+
+    def state_dict(self):
+        """Returns the state of the scheduler as a :class:`dict`.
+
+        It contains an entry for every variable in self.__dict__ which
+        is not the optimizer.
+        """
+        return {
+            "base_lrs": self.base_lrs,
+            "epoch": self.epoch,
+            "batch": self.batch,
+        }
+
+    def load_state_dict(self, state_dict):
+        """Loads the schedulers state.
+
+        Args:
+            state_dict (dict): scheduler state. Should be an object returned
+                from a call to :meth:`state_dict`.
+        """
+        self.__dict__.update(state_dict)
+
+    def get_last_lr(self) -> List[float]:
+        """Return last computed learning rate by current scheduler.  Will be a list of float."""
+        return self._last_lr
+
+    def get_lr(self):
+        # Compute list of learning rates from self.epoch and self.batch and
+        # self.base_lrs; this must be overloaded by the user.
+        # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
+        raise NotImplementedError
+
+    def step_batch(self, batch: Optional[int] = None) -> None:
+        # Step the batch index, or just set it.  If `batch` is specified, it
+        # must be the batch index from the start of training, i.e. summed over
+        # all epochs.
+        # You can call this in any order; if you don't provide 'batch', it should
+        # of course be called once per batch.
+        if batch is not None:
+            self.batch = batch
+        else:
+            self.batch = self.batch + 1
+        self._set_lrs()
+
+    def step_epoch(self, epoch: Optional[Union[int, float]] = None):
+        # Step the epoch index, or just set it.  If you provide the 'epoch' arg,
+        # you should call this at the start of the epoch; if you don't provide the 'epoch'
+        # arg, you should call it at the end of the epoch.
+        if epoch is not None:
+            self.epoch = epoch
+        else:
+            self.epoch = self.epoch + 1
+        self._set_lrs()
+
+    def _set_lrs(self):
+        values = self.get_lr()
+        assert len(values) == len(self.optimizer.param_groups)
+
+        for i, data in enumerate(zip(self.optimizer.param_groups, values)):
+            param_group, lr = data
+            param_group["lr"] = lr
+            self.print_lr(self.verbose, i, lr)
+        self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
+
+    def print_lr(self, is_verbose, group, lr):
+        """Display the current learning rate."""
+        if is_verbose:
+            logging.info(
+                f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
+                f" of group {group} to {lr:.4e}."
+            )
+
+
+class Eden(LRScheduler):
+    """
+    Eden scheduler.
+    The basic formula (before warmup) is:
+      lr = base_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
+                     (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25)) * warmup
+    where `warmup` increases from linearly 0.5 to 1 over `warmup_batches` batches
+    and then stays constant at 1.
+
+
+     E.g. suggest base_lr = 0.04 (passed to optimizer) if used with ScaledAdam
+
+    Args:
+        optimizer: the optimizer to change the learning rates on
+        lr_batches: the number of batches after which we start significantly
+              decreasing the learning rate, suggest 5000.
+        lr_epochs: the number of epochs after which we start significantly
+              decreasing the learning rate, suggest 6 if you plan to do e.g.
+              20 to 40 epochs, but may need smaller number if dataset is huge
+              and you will do few epochs.
+    """
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        lr_batches: Union[int, float],
+        lr_epochs: Union[int, float],
+        warmup_batches: Union[int, float] = 500.0,
+        verbose: bool = False,
+    ):
+        super(Eden, self).__init__(optimizer, verbose)
+        self.lr_batches = lr_batches
+        self.lr_epochs = lr_epochs
+        self.warmup_batches = warmup_batches
+
+    def get_lr(self):
+        factor = (
+            (self.batch**2 + self.lr_batches**2) / self.lr_batches**2
+        ) ** -0.25 * (
+            ((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25
+        )
+        warmup_factor = (
+            1.0
+            if self.batch >= self.warmup_batches
+            else 0.5 + 0.5 * (self.batch / self.warmup_batches)
+        )
+
+        return [x * factor * warmup_factor for x in self.base_lrs]
+
+
+def _test_eden():
+    m = torch.nn.Linear(100, 100)
+    optim = ScaledAdam(m.parameters(), lr=0.03)
+
+    scheduler = Eden(optim, lr_batches=100, lr_epochs=2, verbose=True)
+
+    for epoch in range(10):
+        scheduler.step_epoch(epoch)  # sets epoch to `epoch`
+
+        for step in range(20):
+            x = torch.randn(200, 100).detach()
+            x.requires_grad = True
+            y = m(x)
+            dy = torch.randn(200, 100).detach()
+            f = (y * dy).sum()
+            f.backward()
+
+            optim.step()
+            scheduler.step_batch()
+            optim.zero_grad()
+
+    logging.info(f"last lr = {scheduler.get_last_lr()}")
+    logging.info(f"state dict = {scheduler.state_dict()}")
+
+
+# This is included mostly as a baseline for ScaledAdam.
+class Eve(Optimizer):
+    """
+    Implements Eve algorithm.  This is a modified version of AdamW with a special
+    way of setting the weight-decay / shrinkage-factor, which is designed to make the
+    rms of the parameters approach a particular target_rms (default: 0.1).  This is
+    for use with networks with 'scaled' versions of modules (see scaling.py), which
+    will be close to invariant to the absolute scale on the parameter matrix.
+
+    The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
+    The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
+    Eve is unpublished so far.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float, optional): learning rate (default: 1e-3)
+        betas (Tuple[float, float], optional): coefficients used for computing
+            running averages of gradient and its square (default: (0.9, 0.999))
+        eps (float, optional): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float, optional): weight decay coefficient (default: 3e-4;
+            this value means that the weight would decay significantly after
+            about 3k minibatches.  Is not multiplied by learning rate, but
+            is conditional on RMS-value of parameter being > target_rms.
+        target_rms (float, optional): target root-mean-square value of
+           parameters, if they fall below this we will stop applying weight decay.
+
+
+    .. _Adam: A Method for Stochastic Optimization:
+        https://arxiv.org/abs/1412.6980
+    .. _Decoupled Weight Decay Regularization:
+        https://arxiv.org/abs/1711.05101
+    .. _On the Convergence of Adam and Beyond:
+        https://openreview.net/forum?id=ryQu7f-RZ
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=1e-3,
+        betas=(0.9, 0.98),
+        eps=1e-8,
+        weight_decay=1e-3,
+        target_rms=0.1,
+    ):
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
+        if not 0 <= weight_decay <= 0.1:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0 < target_rms <= 10.0:
+            raise ValueError("Invalid target_rms value: {}".format(target_rms))
+        defaults = dict(
+            lr=lr,
+            betas=betas,
+            eps=eps,
+            weight_decay=weight_decay,
+            target_rms=target_rms,
+        )
+        super(Eve, self).__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super(Eve, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+
+                # Perform optimization step
+                grad = p.grad
+                if grad.is_sparse:
+                    raise RuntimeError("AdamW does not support sparse gradients")
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state["step"] = 0
+                    # Exponential moving average of gradient values
+                    state["exp_avg"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+                    # Exponential moving average of squared gradient values
+                    state["exp_avg_sq"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+
+                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
+
+                beta1, beta2 = group["betas"]
+
+                state["step"] += 1
+                bias_correction1 = 1 - beta1 ** state["step"]
+                bias_correction2 = 1 - beta2 ** state["step"]
+
+                # Decay the first and second moment running average coefficient
+                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
+                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+                denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_(
+                    group["eps"]
+                )
+
+                step_size = group["lr"] / bias_correction1
+                target_rms = group["target_rms"]
+                weight_decay = group["weight_decay"]
+
+                if p.numel() > 1:
+                    # avoid applying this weight-decay on "scaling factors"
+                    # (which are scalar).
+                    is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5))
+                    p.mul_(1 - (weight_decay * is_above_target_rms))
+
+                p.addcdiv_(exp_avg, denom, value=-step_size)
+
+                if random.random() < 0.0005:
+                    step = (exp_avg / denom) * step_size
+                    logging.info(
+                        f"Delta rms = {(step**2).mean().item()}, shape = {step.shape}"
+                    )
+
+        return loss
+
+
+def _test_scaled_adam(hidden_dim: int):
+    import timeit
+
+    from scaling import ScaledLinear
+
+    E = 100
+    B = 4
+    T = 2
+    logging.info("in test_eve_cain")
+    # device = torch.device('cuda')
+    device = torch.device("cpu")
+    dtype = torch.float32
+
+    fix_random_seed(42)
+    # these input_magnitudes and output_magnitudes are to test that
+    # Abel is working as we expect and is able to adjust scales of
+    # different dims differently.
+    input_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp()
+    output_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp()
+
+    for iter in [1, 0]:
+        fix_random_seed(42)
+        Linear = torch.nn.Linear if iter == 0 else ScaledLinear
+
+        m = torch.nn.Sequential(
+            Linear(E, hidden_dim),
+            torch.nn.PReLU(),
+            Linear(hidden_dim, hidden_dim),
+            torch.nn.PReLU(),
+            Linear(hidden_dim, E),
+        ).to(device)
+
+        train_pairs = [
+            (
+                100.0
+                * torch.randn(B, T, E, device=device, dtype=dtype)
+                * input_magnitudes,
+                torch.randn(B, T, E, device=device, dtype=dtype) * output_magnitudes,
+            )
+            for _ in range(20)
+        ]
+
+        if iter == 0:
+            optim = Eve(m.parameters(), lr=0.003)
+        elif iter == 1:
+            optim = ScaledAdam(m.parameters(), lr=0.03, clipping_scale=2.0)
+        scheduler = Eden(optim, lr_batches=200, lr_epochs=5, verbose=False)
+
+        start = timeit.default_timer()
+        avg_loss = 0.0
+        for epoch in range(180):
+            scheduler.step_epoch()
+            # if epoch == 100 and iter in [2,3]:
+            #    optim.reset_speedup()  # check it doesn't crash.
+
+            # if epoch == 130:
+            #    opts = diagnostics.TensorDiagnosticOptions(
+            #        2 ** 22
+            #    )  # allow 4 megabytes per sub-module
+            #    diagnostic = diagnostics.attach_diagnostics(m, opts)
+
+            for n, (x, y) in enumerate(train_pairs):
+                y_out = m(x)
+                loss = ((y_out - y) ** 2).mean() * 100.0
+                if epoch == 0 and n == 0:
+                    avg_loss = loss.item()
+                else:
+                    avg_loss = 0.98 * avg_loss + 0.02 * loss.item()
+                if n == 0 and epoch % 5 == 0:
+                    # norm1 = '%.2e' % (m[0].weight**2).mean().sqrt().item()
+                    # norm1b = '%.2e' % (m[0].bias**2).mean().sqrt().item()
+                    # norm2 = '%.2e' % (m[2].weight**2).mean().sqrt().item()
+                    # norm2b = '%.2e' % (m[2].bias**2).mean().sqrt().item()
+                    # scale1 = '%.2e' % (m[0].weight_scale.exp().item())
+                    # scale1b = '%.2e' % (m[0].bias_scale.exp().item())
+                    # scale2 = '%.2e' % (m[2].weight_scale.exp().item())
+                    # scale2b = '%.2e' % (m[2].bias_scale.exp().item())
+                    lr = scheduler.get_last_lr()[0]
+                    logging.info(
+                        f"Iter {iter}, epoch {epoch}, batch {n}, avg_loss {avg_loss:.4g}, lr={lr:.4e}"
+                    )  # , norms={norm1,norm1b,norm2,norm2b}") # scales={scale1,scale1b,scale2,scale2b}
+                loss.log().backward()
+                optim.step()
+                optim.zero_grad()
+                scheduler.step_batch()
+
+        # diagnostic.print_diagnostics()
+
+        stop = timeit.default_timer()
+        logging.info(f"Iter={iter}, Time taken: {stop - start}")
+
+        logging.info(f"last lr = {scheduler.get_last_lr()}")
+        # logging.info("state dict = ", scheduler.state_dict())
+        # logging.info("optim state_dict = ", optim.state_dict())
+        logging.info(f"input_magnitudes = {input_magnitudes}")
+        logging.info(f"output_magnitudes = {output_magnitudes}")
+
+
+if __name__ == "__main__":
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    logging.getLogger().setLevel(logging.INFO)
+    import subprocess
+
+    s = subprocess.check_output(
+        "git status -uno .; git log -1; git diff HEAD .", shell=True
+    )
+    logging.info(s)
+    import sys
+
+    if len(sys.argv) > 1:
+        hidden_dim = int(sys.argv[1])
+    else:
+        hidden_dim = 200
+
+    _test_scaled_adam(hidden_dim)
+    _test_eden()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/pretrained.py b/egs/libriheavy/ASR/zipformer_prompt_asr/pretrained.py
new file mode 100644
index 000000000..48fd2612a
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/pretrained.py
@@ -0,0 +1,359 @@
+#!/usr/bin/env python3
+# Copyright      2021-2023  Xiaomi Corp.        (authors: Fangjun Kuang, Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script loads a checkpoint (`pretrained.pt`) and uses it to decode waves.
+You can generate the checkpoint with the following command:
+
+./zipformer/export_PromptASR.py \
+  --exp-dir ./zipformer/exp \
+  --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \
+  --epoch 50 \
+  --avg 10
+
+Utterance level context biasing:
+
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \
+  --method modified_beam_search \
+  --use-pre-text True \
+  --content-prompt "bessy random words hello k2 ASR" \
+  --use-style-prompt True \
+  librispeech.flac
+
+
+Word level context biasing:
+
+./zipformer/pretrained.py \
+  --checkpoint ./zipformer/exp/pretrained.pt \
+  --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \
+  --method modified_beam_search \
+  --use-pre-text True \
+  --content-prompt "The topic is about horses." \
+  --use-style-prompt True \
+  test.wav
+
+
+"""
+
+import argparse
+import logging
+import math
+import warnings
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import greedy_search_batch, modified_beam_search
+from text_normalization import _apply_style_transform, train_text_normalization
+from torch.nn.utils.rnn import pad_sequence
+from train_bert_encoder import (
+    _encode_texts_as_bytes_with_tokenizer,
+    add_model_arguments,
+    get_params,
+    get_tokenizer,
+    get_transducer_model,
+)
+
+from icefall.utils import make_pad_mask, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500_fallback_coverage_0.99/bpe.model",
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-pre-text",
+        type=str2bool,
+        default=True,
+        help="Use content prompt during decoding",
+    )
+
+    parser.add_argument(
+        "--use-style-prompt",
+        type=str2bool,
+        default=True,
+        help="Use style prompt during decoding",
+    )
+
+    parser.add_argument(
+        "--pre-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of content prompt, i.e pre_text",
+    )
+
+    parser.add_argument(
+        "--style-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of style prompt, i.e style_text",
+    )
+
+    parser.add_argument(
+        "--content-prompt", type=str, default="", help="The content prompt for decoding"
+    )
+
+    parser.add_argument(
+        "--style-prompt",
+        type=str,
+        default="Mixed-cased English text with punctuations, feel free to change it.",
+        help="The style prompt for decoding",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    if params.causal:
+        assert (
+            "," not in params.chunk_size
+        ), "chunk_size should be one value in decoding."
+        assert (
+            "," not in params.left_context_frames
+        ), "left_context_frames should be one value in decoding."
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+    tokenizer = get_tokenizer(params)  # for text encoder
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    assert (
+        len(params.sound_files) == 1
+    ), "Only support decoding one audio at this moment"
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    # encode prompts
+    if params.use_pre_text:
+        pre_text = [train_text_normalization(params.content_prompt)]
+        pre_text = _apply_style_transform(pre_text, params.pre_text_transform)
+    else:
+        pre_text = [""]
+
+    if params.use_style_prompt:
+        style_text = [params.style_prompt]
+        style_text = _apply_style_transform(style_text, params.style_text_transform)
+    else:
+        style_text = [""]
+
+    if params.use_pre_text or params.use_style_prompt:
+        encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer(
+            pre_texts=pre_text,
+            style_texts=style_text,
+            tokenizer=tokenizer,
+            device=device,
+            no_limit=True,
+        )
+
+        memory, memory_key_padding_mask = model.encode_text(
+            encoded_inputs=encoded_inputs,
+            style_lens=style_lens,
+        )  # (T,B,C)
+    else:
+        memory = None
+        memory_key_padding_mask = None
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        encoder_out, encoder_out_lens = model.encode_audio(
+            feature=features,
+            feature_lens=feature_lengths,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )
+
+    hyps = []
+    msg = f"Using {params.method}"
+    logging.info(msg)
+
+    if params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+        hyps.append(sp.decode(hyp_tokens)[0])
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        hyps.append(sp.decode(hyp_tokens)[0])
+    else:
+        raise ValueError(f"Unsupported method: {params.method}")
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        s += f"{filename}:\n{hyp}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/scaling.py b/egs/libriheavy/ASR/zipformer_prompt_asr/scaling.py
new file mode 100644
index 000000000..0e6764ba0
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/scaling.py
@@ -0,0 +1,1872 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import collections
+import logging
+import math
+import random
+from functools import reduce
+from itertools import repeat
+from typing import Optional, Tuple, Union
+
+import k2
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import Tensor
+from torch.cuda.amp import custom_bwd, custom_fwd
+from torch.nn import Embedding as ScaledEmbedding
+
+
+class PiecewiseLinear(object):
+    """
+    Piecewise linear function, from float to float, specified as nonempty list of (x,y) pairs with
+    the x values in order.  x values <[initial x] or >[final x] are map to [initial y], [final y]
+    respectively.
+    """
+
+    def __init__(self, *args):
+        assert len(args) >= 1
+        if len(args) == 1 and isinstance(args[0], PiecewiseLinear):
+            self.pairs = list(args[0].pairs)
+        else:
+            self.pairs = [(float(x), float(y)) for x, y in args]
+        for (x, y) in self.pairs:
+            assert isinstance(x, float) or isinstance(x, int)
+            assert isinstance(y, float) or isinstance(y, int)
+
+        for i in range(len(self.pairs) - 1):
+            assert self.pairs[i + 1][0] > self.pairs[i][0], self.pairs
+
+    def __str__(self):
+        # e.g. 'PiecewiseLinear((0., 10.), (100., 0.))'
+        return f"PiecewiseLinear({str(self.pairs)[1:-1]})"
+
+    def __call__(self, x):
+        if x <= self.pairs[0][0]:
+            return self.pairs[0][1]
+        elif x >= self.pairs[-1][0]:
+            return self.pairs[-1][1]
+        else:
+            cur_x, cur_y = self.pairs[0]
+            for i in range(1, len(self.pairs)):
+                next_x, next_y = self.pairs[i]
+                if x >= cur_x and x <= next_x:
+                    return cur_y + (next_y - cur_y) * (x - cur_x) / (next_x - cur_x)
+                cur_x, cur_y = next_x, next_y
+            assert False
+
+    def __mul__(self, alpha):
+        return PiecewiseLinear(*[(x, y * alpha) for x, y in self.pairs])
+
+    def __add__(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            return PiecewiseLinear(*[(p[0], p[1] + x) for p in self.pairs])
+        s, x = self.get_common_basis(x)
+        return PiecewiseLinear(
+            *[(sp[0], sp[1] + xp[1]) for sp, xp in zip(s.pairs, x.pairs)]
+        )
+
+    def max(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            x = PiecewiseLinear((0, x))
+        s, x = self.get_common_basis(x, include_crossings=True)
+        return PiecewiseLinear(
+            *[(sp[0], max(sp[1], xp[1])) for sp, xp in zip(s.pairs, x.pairs)]
+        )
+
+    def min(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            x = PiecewiseLinear((0, x))
+        s, x = self.get_common_basis(x, include_crossings=True)
+        return PiecewiseLinear(
+            *[(sp[0], min(sp[1], xp[1])) for sp, xp in zip(s.pairs, x.pairs)]
+        )
+
+    def __eq__(self, other):
+        return self.pairs == other.pairs
+
+    def get_common_basis(self, p: "PiecewiseLinear", include_crossings: bool = False):
+        """
+        Returns (self_mod, p_mod) which are equivalent piecewise lienar
+        functions to self and p, but with the same x values.
+
+          p: the other piecewise linear function
+          include_crossings: if true, include in the x values positions
+              where the functions indicate by this and p crosss.
+        """
+        assert isinstance(p, PiecewiseLinear)
+
+        # get sorted x-values without repetition.
+        x_vals = sorted(set([x for x, y in self.pairs] + [x for x, y in p.pairs]))
+        y_vals1 = [self(x) for x in x_vals]
+        y_vals2 = [p(x) for x in x_vals]
+
+        if include_crossings:
+            extra_x_vals = []
+            for i in range(len(x_vals) - 1):
+                if (y_vals1[i] > y_vals2[i]) != (y_vals1[i + 1] > y_vals2[i + 1]):
+                    # if the two lines in this subsegment potentially cross each other..
+                    diff_cur = abs(y_vals1[i] - y_vals2[i])
+                    diff_next = abs(y_vals1[i + 1] - y_vals2[i + 1])
+                    # `pos`, between 0 and 1, gives the relative x position,
+                    # with 0 being x_vals[i] and 1 being x_vals[i+1].
+                    pos = diff_cur / (diff_cur + diff_next)
+                    extra_x_val = x_vals[i] + pos * (x_vals[i + 1] - x_vals[i])
+                    extra_x_vals.append(extra_x_val)
+            if len(extra_x_vals) > 0:
+                x_vals = sorted(set(x_vals + extra_x_vals))
+        y_vals1 = [self(x) for x in x_vals]
+        y_vals2 = [p(x) for x in x_vals]
+        return (
+            PiecewiseLinear(*zip(x_vals, y_vals1)),
+            PiecewiseLinear(*zip(x_vals, y_vals2)),
+        )
+
+
+class ScheduledFloat(torch.nn.Module):
+    """
+    This object is a torch.nn.Module only because we want it to show up in [top_level module].modules();
+    it does not have a working forward() function.  You are supposed to cast it to float, as
+    in, float(parent_module.whatever), and use it as something like a dropout prob.
+
+    It is a floating point value whose value changes depending on the batch count of the
+    training loop.  It is a piecewise linear function where you specifiy the (x,y) pairs
+    in sorted order on x; x corresponds to the batch index.  For batch-index values before the
+    first x or after the last x, we just use the first or last y value.
+
+    Example:
+       self.dropout = ScheduledFloat((0.0, 0.2), (4000.0, 0.0), default=0.0)
+
+    `default` is used when self.batch_count is not set or in training or mode or in
+     torch.jit scripting mode.
+    """
+
+    def __init__(self, *args, default: float = 0.0):
+        super().__init__()
+        # self.batch_count and self.name will be written to in the training loop.
+        self.batch_count = None
+        self.name = None
+        self.default = default
+        self.schedule = PiecewiseLinear(*args)
+
+    def extra_repr(self) -> str:
+        return (
+            f"batch_count={self.batch_count}, schedule={str(self.schedule.pairs[1:-1])}"
+        )
+
+    def __float__(self):
+        batch_count = self.batch_count
+        if batch_count is None or not self.training or torch.jit.is_scripting():
+            return float(self.default)
+        else:
+            ans = self.schedule(self.batch_count)
+            if random.random() < 0.0002:
+                logging.info(
+                    f"ScheduledFloat: name={self.name}, batch_count={self.batch_count}, ans={ans}"
+                )
+            return ans
+
+    def __add__(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            return ScheduledFloat(self.schedule + x, default=self.default)
+        else:
+            return ScheduledFloat(
+                self.schedule + x.schedule, default=self.default + x.default
+            )
+
+    def max(self, x):
+        if isinstance(x, float) or isinstance(x, int):
+            return ScheduledFloat(self.schedule.max(x), default=self.default)
+        else:
+            return ScheduledFloat(
+                self.schedule.max(x.schedule), default=max(self.default, x.default)
+            )
+
+
+FloatLike = Union[float, ScheduledFloat]
+
+
+def random_cast_to_half(x: Tensor, min_abs: float = 5.0e-06) -> Tensor:
+    """
+    A randomized way of casting a floating point value to half precision.
+    """
+    if x.dtype == torch.float16:
+        return x
+    x_abs = x.abs()
+    is_too_small = x_abs < min_abs
+    # for elements where is_too_small is true, random_val will contain +-min_abs with
+    # probability (x.abs() / min_abs), and 0.0 otherwise.  [so this preserves expectations,
+    # for those elements].
+    random_val = min_abs * x.sign() * (torch.rand_like(x) * min_abs < x_abs)
+    return torch.where(is_too_small, random_val, x).to(torch.float16)
+
+
+class CutoffEstimator:
+    """
+    Estimates cutoffs of an arbitrary numerical quantity such that a specified
+    proportion of items will be above the cutoff on average.
+
+      p is the proportion of items that should be above the cutoff.
+    """
+
+    def __init__(self, p: float):
+        self.p = p
+        # total count of items
+        self.count = 0
+        # total count of items that were above the cutoff
+        self.count_above = 0
+        # initial cutoff value
+        self.cutoff = 0
+
+    def __call__(self, x: float) -> bool:
+        """
+        Returns true if x is above the cutoff.
+        """
+        ans = x > self.cutoff
+        self.count += 1
+        if ans:
+            self.count_above += 1
+        cur_p = self.count_above / self.count
+        delta_p = cur_p - self.p
+        if (delta_p > 0) == ans:
+            q = abs(delta_p)
+            self.cutoff = x * q + self.cutoff * (1 - q)
+        return ans
+
+
+class SoftmaxFunction(torch.autograd.Function):
+    """
+    Tries to handle half-precision derivatives in a randomized way that should
+    be more accurate for training than the default behavior.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor, dim: int):
+        ans = x.softmax(dim=dim)
+        # if x dtype is float16, x.softmax() returns a float32 because
+        # (presumably) that op does not support float16, and autocast
+        # is enabled.
+        if torch.is_autocast_enabled():
+            ans = ans.to(torch.float16)
+        ctx.save_for_backward(ans)
+        ctx.x_dtype = x.dtype
+        ctx.dim = dim
+        return ans
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor):
+        (ans,) = ctx.saved_tensors
+        with torch.cuda.amp.autocast(enabled=False):
+            ans_grad = ans_grad.to(torch.float32)
+            ans = ans.to(torch.float32)
+            x_grad = ans_grad * ans
+            x_grad = x_grad - ans * x_grad.sum(dim=ctx.dim, keepdim=True)
+            return x_grad, None
+
+
+def softmax(x: Tensor, dim: int):
+    return SoftmaxFunction.apply(x, dim)
+
+
+class MaxEigLimiterFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        coeffs: Tensor,
+        direction: Tensor,
+        channel_dim: int,
+        grad_scale: float,
+    ) -> Tensor:
+        ctx.channel_dim = channel_dim
+        ctx.grad_scale = grad_scale
+        ctx.save_for_backward(x.detach(), coeffs.detach(), direction.detach())
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad, *args):
+        with torch.enable_grad():
+            (x_orig, coeffs, new_direction) = ctx.saved_tensors
+            x_orig.requires_grad = True
+            num_channels = x_orig.shape[ctx.channel_dim]
+            x = x_orig.transpose(ctx.channel_dim, -1).reshape(-1, num_channels)
+            new_direction.requires_grad = False
+            x = x - x.mean(dim=0)
+            x_var = (x**2).mean()
+            x_residual = x - coeffs * new_direction
+            x_residual_var = (x_residual**2).mean()
+            # `variance_proportion` is the proportion of the variance accounted for
+            # by the top eigen-direction.  This is to be minimized.
+            variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20)
+            variance_proportion.backward()
+        x_orig_grad = x_orig.grad
+        x_extra_grad = (
+            x_orig.grad
+            * ctx.grad_scale
+            * x_grad.norm()
+            / (x_orig_grad.norm() + 1.0e-20)
+        )
+        return x_grad + x_extra_grad.detach(), None, None, None, None
+
+
+class BiasNormFunction(torch.autograd.Function):
+    # This computes:
+    #   scales = (torch.mean((x - bias) ** 2, keepdim=True)) ** -0.5 * log_scale.exp()
+    #   return (x - bias) * scales
+    # (after unsqueezing the bias), but it does it in a memory-efficient way so that
+    # it can just store the returned value (chances are, this will also be needed for
+    # some other reason, related to the next operation, so we can save memory).
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        bias: Tensor,
+        log_scale: Tensor,
+        channel_dim: int,
+        store_output_for_backprop: bool,
+    ) -> Tensor:
+        assert bias.ndim == 1
+        if channel_dim < 0:
+            channel_dim = channel_dim + x.ndim
+        ctx.store_output_for_backprop = store_output_for_backprop
+        ctx.channel_dim = channel_dim
+        for _ in range(channel_dim + 1, x.ndim):
+            bias = bias.unsqueeze(-1)
+        scales = (
+            torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) ** -0.5
+        ) * log_scale.exp()
+        ans = x * scales
+        ctx.save_for_backward(
+            ans.detach() if store_output_for_backprop else x,
+            scales.detach(),
+            bias.detach(),
+            log_scale.detach(),
+        )
+        return ans
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor) -> Tensor:
+        ans_or_x, scales, bias, log_scale = ctx.saved_tensors
+        if ctx.store_output_for_backprop:
+            x = ans_or_x / scales
+        else:
+            x = ans_or_x
+        x = x.detach()
+        x.requires_grad = True
+        bias.requires_grad = True
+        log_scale.requires_grad = True
+        with torch.enable_grad():
+            # recompute scales from x, bias and log_scale.
+            scales = (
+                torch.mean((x - bias) ** 2, dim=ctx.channel_dim, keepdim=True) ** -0.5
+            ) * log_scale.exp()
+            ans = x * scales
+            ans.backward(gradient=ans_grad)
+        return x.grad, bias.grad.flatten(), log_scale.grad, None, None
+
+
+class BiasNorm(torch.nn.Module):
+    """
+    This is intended to be a simpler, and hopefully cheaper, replacement for
+    LayerNorm.  The observation this is based on, is that Transformer-type
+    networks, especially with pre-norm, sometimes seem to set one of the
+    feature dimensions to a large constant value (e.g. 50), which "defeats"
+    the LayerNorm because the output magnitude is then not strongly dependent
+    on the other (useful) features.  Presumably the weight and bias of the
+    LayerNorm are required to allow it to do this.
+
+    Instead, we give the BiasNorm a trainable bias that it can use when
+    computing the scale for normalization.  We also give it a (scalar)
+    trainable scale on the output.
+
+
+    Args:
+       num_channels: the number of channels, e.g. 512.
+       channel_dim: the axis/dimension corresponding to the channel,
+         interprted as an offset from the input's ndim if negative.
+         shis is NOT the num_channels; it should typically be one of
+         {-2, -1, 0, 1, 2, 3}.
+      log_scale: the initial log-scale that we multiply the output by; this
+         is learnable.
+      log_scale_min: FloatLike, minimum allowed value of log_scale
+      log_scale_max: FloatLike, maximum allowed value of log_scale
+      store_output_for_backprop: only possibly affects memory use; recommend
+         to set to True if you think the output of this module is more likely
+         than the input of this module to be required to be stored for the
+         backprop.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+        log_scale: float = 1.0,
+        log_scale_min: float = -1.5,
+        log_scale_max: float = 1.5,
+        store_output_for_backprop: bool = False,
+    ) -> None:
+        super(BiasNorm, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.log_scale = nn.Parameter(torch.tensor(log_scale))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+
+        self.log_scale_min = log_scale_min
+        self.log_scale_max = log_scale_max
+
+        self.store_output_for_backprop = store_output_for_backprop
+
+    def forward(self, x: Tensor) -> Tensor:
+        assert x.shape[self.channel_dim] == self.num_channels
+
+        if torch.jit.is_scripting():
+            channel_dim = self.channel_dim
+            if channel_dim < 0:
+                channel_dim += x.ndim
+            bias = self.bias
+            for _ in range(channel_dim + 1, x.ndim):
+                bias = bias.unsqueeze(-1)
+            scales = (
+                torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) ** -0.5
+            ) * self.log_scale.exp()
+            return x * scales
+
+        log_scale = limit_param_value(
+            self.log_scale,
+            min=float(self.log_scale_min),
+            max=float(self.log_scale_max),
+            training=self.training,
+        )
+
+        return BiasNormFunction.apply(
+            x, self.bias, log_scale, self.channel_dim, self.store_output_for_backprop
+        )
+
+
+def ScaledLinear(*args, initial_scale: float = 1.0, **kwargs) -> nn.Linear:
+    """
+    Behaves like a constructor of a modified version of nn.Linear
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Linear(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+def ScaledConv1d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv1d:
+    """
+    Behaves like a constructor of a modified version of nn.Conv1d
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Conv1d(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+def ScaledConv2d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv2d:
+    """
+    Behaves like a constructor of a modified version of nn.Conv2d
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False, but:
+    NO PADDING-RELATED ARGS.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Conv2d(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
+    return ans
+
+
+class ChunkCausalDepthwiseConv1d(torch.nn.Module):
+    """
+    Behaves like a depthwise 1d convolution, except that it is causal in
+    a chunkwise way, as if we had a block-triangular attention mask.
+    The chunk size is provided at test time (it should probably be
+    kept in sync with the attention mask).
+
+    This has a little more than twice the parameters of a conventional
+    depthwise conv1d module: we implement it by having one
+    depthwise convolution, of half the width, that is causal (via
+    right-padding); and one depthwise convolution that is applied only
+    within chunks, that we multiply by a scaling factor which depends
+    on the position within the chunk.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        initial_scale: float = 1.0,
+        bias: bool = True,
+    ):
+        super().__init__()
+        assert kernel_size % 2 == 1
+
+        half_kernel_size = (kernel_size + 1) // 2
+        # will pad manually, on one side.
+        self.causal_conv = nn.Conv1d(
+            in_channels=channels,
+            out_channels=channels,
+            groups=channels,
+            kernel_size=half_kernel_size,
+            padding=0,
+            bias=True,
+        )
+
+        self.chunkwise_conv = nn.Conv1d(
+            in_channels=channels,
+            out_channels=channels,
+            groups=channels,
+            kernel_size=kernel_size,
+            padding=kernel_size // 2,
+            bias=bias,
+        )
+
+        # first row is correction factors added to the scale near the left edge of the chunk,
+        # second row is correction factors added to the scale near the right edge of the chunk,
+        # both of these are added to a default scale of 1.0.
+        self.chunkwise_conv_scale = nn.Parameter(torch.zeros(2, channels, kernel_size))
+        self.kernel_size = kernel_size
+
+        with torch.no_grad():
+            self.causal_conv.weight[:] *= initial_scale
+            self.chunkwise_conv.weight[:] *= initial_scale
+            if bias:
+                torch.nn.init.uniform_(
+                    self.causal_conv.bias, -0.1 * initial_scale, 0.1 * initial_scale
+                )
+
+    def forward(self, x: Tensor, chunk_size: int = -1) -> Tensor:
+        """
+             Forward function.  Args:
+               x: a Tensor of shape (batch_size, channels, seq_len)
+        chunk_size: the chunk size, in frames; does not have to divide seq_len exactly.
+        """
+        (batch_size, num_channels, seq_len) = x.shape
+
+        half_kernel_size = self.kernel_size + 1 // 2
+        # left_pad is half_kernel_size - 1 where half_kernel_size is the size used
+        # in the causal conv.  It's the amount by which we must pad on the left,
+        # to make the convolution causal.
+        left_pad = self.kernel_size // 2
+
+        if chunk_size < 0 or chunk_size > seq_len:
+            chunk_size = seq_len
+        right_pad = -seq_len % chunk_size
+
+        x = torch.nn.functional.pad(x, (left_pad, right_pad))
+
+        x_causal = self.causal_conv(x[..., : left_pad + seq_len])
+        assert x_causal.shape == (batch_size, num_channels, seq_len)
+
+        x_chunk = x[..., left_pad:]
+        num_chunks = x_chunk.shape[2] // chunk_size
+        x_chunk = x_chunk.reshape(batch_size, num_channels, num_chunks, chunk_size)
+        x_chunk = x_chunk.permute(0, 2, 1, 3).reshape(
+            batch_size * num_chunks, num_channels, chunk_size
+        )
+        x_chunk = self.chunkwise_conv(x_chunk)  # does not change shape
+
+        chunk_scale = self._get_chunk_scale(chunk_size)
+
+        x_chunk = x_chunk * chunk_scale
+        x_chunk = x_chunk.reshape(
+            batch_size, num_chunks, num_channels, chunk_size
+        ).permute(0, 2, 1, 3)
+        x_chunk = x_chunk.reshape(batch_size, num_channels, num_chunks * chunk_size)[
+            ..., :seq_len
+        ]
+
+        return x_chunk + x_causal
+
+    def _get_chunk_scale(self, chunk_size: int):
+        """Returns tensor of shape (num_channels, chunk_size) that will be used to
+        scale the output of self.chunkwise_conv."""
+        left_edge = self.chunkwise_conv_scale[0]
+        right_edge = self.chunkwise_conv_scale[1]
+        if chunk_size < self.kernel_size:
+            left_edge = left_edge[:, :chunk_size]
+            right_edge = right_edge[:, -chunk_size:]
+        else:
+            t = chunk_size - self.kernel_size
+            channels = left_edge.shape[0]
+            pad = torch.zeros(
+                channels, t, device=left_edge.device, dtype=left_edge.dtype
+            )
+            left_edge = torch.cat((left_edge, pad), dim=-1)
+            right_edge = torch.cat((pad, right_edge), dim=-1)
+        return 1.0 + (left_edge + right_edge)
+
+
+class BalancerFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        min_mean: float,
+        max_mean: float,
+        min_rms: float,
+        max_rms: float,
+        grad_scale: float,
+        channel_dim: int,
+    ) -> Tensor:
+        if channel_dim < 0:
+            channel_dim += x.ndim
+        ctx.channel_dim = channel_dim
+        ctx.save_for_backward(x)
+        ctx.config = (min_mean, max_mean, min_rms, max_rms, grad_scale, channel_dim)
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None, None, None]:
+        (x,) = ctx.saved_tensors
+        (min_mean, max_mean, min_rms, max_rms, grad_scale, channel_dim) = ctx.config
+
+        try:
+            with torch.enable_grad():
+                with torch.cuda.amp.autocast(enabled=False):
+                    x = x.to(torch.float32)
+                    x = x.detach()
+                    x.requires_grad = True
+                    mean_dims = [i for i in range(x.ndim) if i != channel_dim]
+                    uncentered_var = (x**2).mean(dim=mean_dims, keepdim=True)
+                    mean = x.mean(dim=mean_dims, keepdim=True)
+                    stddev = (uncentered_var - (mean * mean)).clamp(min=1.0e-20).sqrt()
+                    rms = uncentered_var.clamp(min=1.0e-20).sqrt()
+
+                    m = mean / stddev
+                    # part of loss that relates to mean / stddev
+                    m_loss = (m - m.clamp(min=min_mean, max=max_mean)).abs()
+
+                    # put a much larger scale on the RMS-max-limit loss, so that if both it and the
+                    # m_loss are violated we fix the RMS loss first.
+                    rms_clamped = rms.clamp(min=min_rms, max=max_rms)
+                    r_loss = (rms_clamped / rms).log().abs()
+
+                    loss = m_loss + r_loss
+
+                    loss.backward(gradient=torch.ones_like(loss))
+                    loss_grad = x.grad
+                    loss_grad_rms = (
+                        (loss_grad**2)
+                        .mean(dim=mean_dims, keepdim=True)
+                        .sqrt()
+                        .clamp(min=1.0e-20)
+                    )
+
+                    loss_grad = loss_grad * (grad_scale / loss_grad_rms)
+
+                    x_grad_float = x_grad.to(torch.float32)
+                    # scale each element of loss_grad by the absolute value of the corresponding
+                    # element of x_grad, which we view as a noisy estimate of its magnitude for that
+                    # (frame and dimension).  later we can consider factored versions.
+                    x_grad_mod = x_grad_float + (x_grad_float.abs() * loss_grad)
+                    x_grad = x_grad_mod.to(x_grad.dtype)
+        except Exception as e:
+            logging.info(
+                f"Caught exception in Balancer backward: {e}, size={list(x_grad.shape)}, will continue."
+            )
+
+        return x_grad, None, None, None, None, None, None
+
+
+class Balancer(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to encourage, for
+    each channel, that it is positive at least a proportion `threshold` of the
+    time.  It does this by multiplying negative derivative values by up to
+    (1+max_factor), and positive derivative values by up to (1-max_factor),
+    interpolated from 1 at the threshold to those extremal values when none
+    of the inputs are positive.
+
+    Args:
+           num_channels: the number of channels
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           min_positive: the minimum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_positive: the maximum, per channel, of the proportion of the time
+               that (x > 0), above which we start to modify the derivatives.
+           scale_gain_factor: determines the 'gain' with which we increase the
+              change in gradient once the constraints on min_abs and max_abs
+              are violated.
+           min_abs:  the minimum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+           max_abs:  the maximum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+         prob: determines the minimum probability with which we modify the
+             gradients for the {min,max}_positive and {min,max}_abs constraints,
+             on each forward().  This is done randomly to prevent all layers
+             from doing it at the same time.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int,
+        min_positive: FloatLike = 0.05,
+        max_positive: FloatLike = 0.95,
+        min_abs: FloatLike = 0.2,
+        max_abs: FloatLike = 100.0,
+        grad_scale: FloatLike = 0.04,
+        prob: Optional[FloatLike] = None,
+    ):
+        super().__init__()
+
+        if prob is None:
+            prob = ScheduledFloat((0.0, 0.5), (8000.0, 0.125), default=0.4)
+        self.prob = prob
+        # 5% of the time we will return and do nothing because memory usage is
+        # too high.
+        self.mem_cutoff = CutoffEstimator(0.05)
+
+        # actually self.num_channels is no longer needed except for an assertion.
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.min_positive = min_positive
+        self.max_positive = max_positive
+        self.min_abs = min_abs
+        self.max_abs = max_abs
+        self.grad_scale = grad_scale
+
+    def forward(self, x: Tensor) -> Tensor:
+        if (
+            torch.jit.is_scripting()
+            or not x.requires_grad
+            or (x.is_cuda and self.mem_cutoff(torch.cuda.memory_allocated()))
+        ):
+            return _no_op(x)
+
+        prob = float(self.prob)
+        if random.random() < prob:
+            # The following inner-functions convert from the way we historically specified
+            # these limitations, as limits on the absolute value and the proportion of positive
+            # values, to limits on the RMS value and the (mean / stddev).
+            def _abs_to_rms(x):
+                # for normally distributed data, if the expected absolute value is x, the
+                # expected rms value will be sqrt(pi/2) * x.
+                return 1.25331413732 * x
+
+            def _proportion_positive_to_mean(x):
+                def _atanh(x):
+                    eps = 1.0e-10
+                    # eps is to prevent crashes if x is exactly 0 or 1.
+                    # we'll just end up returning a fairly large value.
+                    return (math.log(1 + x + eps) - math.log(1 - x + eps)) / 2.0
+
+                def _approx_inverse_erf(x):
+                    # 1 / (sqrt(pi) * ln(2)),
+                    # see https://math.stackexchange.com/questions/321569/approximating-the-error-function-erf-by-analytical-functions
+                    # this approximation is extremely crude and gets progressively worse for
+                    # x very close to -1 or +1, but we mostly care about the "middle" region
+                    # e.g. _approx_inverse_erf(0.05) = 0.0407316414078772,
+                    # and math.erf(0.0407316414078772) = 0.045935330944660666,
+                    # which is pretty close to 0.05.
+                    return 0.8139535143 * _atanh(x)
+
+                # first convert x from the range 0..1 to the range -1..1 which the error
+                # function returns
+                x = -1 + (2 * x)
+                return _approx_inverse_erf(x)
+
+            min_mean = _proportion_positive_to_mean(float(self.min_positive))
+            max_mean = _proportion_positive_to_mean(float(self.max_positive))
+            min_rms = _abs_to_rms(float(self.min_abs))
+            max_rms = _abs_to_rms(float(self.max_abs))
+            grad_scale = float(self.grad_scale)
+
+            assert x.shape[self.channel_dim] == self.num_channels
+
+            return BalancerFunction.apply(
+                x, min_mean, max_mean, min_rms, max_rms, grad_scale, self.channel_dim
+            )
+        else:
+            return _no_op(x)
+
+
+def penalize_abs_values_gt(
+    x: Tensor, limit: float, penalty: float, name: str = None
+) -> Tensor:
+    """
+    Returns x unmodified, but in backprop will put a penalty for the excess of
+    the absolute values of elements of x over the limit "limit".  E.g. if
+    limit == 10.0, then if x has any values over 10 it will get a penalty.
+
+    Caution: the value of this penalty will be affected by grad scaling used
+    in automatic mixed precision training.  For this reasons we use this,
+    it shouldn't really matter, or may even be helpful; we just use this
+    to disallow really implausible values of scores to be given to softmax.
+
+    The name is for randomly printed debug info.
+    """
+    x_sign = x.sign()
+    over_limit = (x.abs() - limit) > 0
+    # The following is a memory efficient way to penalize the absolute values of
+    # x that's over the limit.  (The memory efficiency comes when you think
+    # about which items torch needs to cache for the autograd, and which ones it
+    # can throw away).  The numerical value of aux_loss as computed here will
+    # actually be larger than it should be, by limit * over_limit.sum(), but it
+    # has the same derivative as the real aux_loss which is penalty * (x.abs() -
+    # limit).relu().
+    aux_loss = penalty * ((x_sign * over_limit).to(torch.int8) * x)
+    # note: we don't do sum() here on aux)_loss, but it's as if we had done
+    # sum() due to how with_loss() works.
+    x = with_loss(x, aux_loss, name)
+    # you must use x for something, or this will be ineffective.
+    return x
+
+
+def _diag(x: Tensor):  # like .diag(), but works for tensors with 3 dims.
+    if x.ndim == 2:
+        return x.diag()
+    else:
+        (batch, dim, dim) = x.shape
+        x = x.reshape(batch, dim * dim)
+        x = x[:, :: dim + 1]
+        assert x.shape == (batch, dim)
+        return x
+
+
+def _whitening_metric(x: Tensor, num_groups: int):
+    """
+    Computes the "whitening metric", a value which will be 1.0 if all the eigenvalues of
+    of the centered feature covariance are the same within each group's covariance matrix
+    and also between groups.
+    Args:
+        x: a Tensor of shape (*, num_channels)
+     num_groups:  the number of groups of channels, a number >=1 that divides num_channels
+    Returns:
+        Returns a scalar Tensor that will be 1.0 if the data is "perfectly white" and
+    greater than 1.0 otherwise.
+    """
+    assert x.dtype != torch.float16
+    x = x.reshape(-1, x.shape[-1])
+    (num_frames, num_channels) = x.shape
+    assert num_channels % num_groups == 0
+    channels_per_group = num_channels // num_groups
+    x = x.reshape(num_frames, num_groups, channels_per_group).transpose(0, 1)
+    # x now has shape (num_groups, num_frames, channels_per_group)
+    # subtract the mean so we use the centered, not uncentered, covariance.
+    # My experience has been that when we "mess with the gradients" like this,
+    # it's better not do anything that tries to move the mean around, because
+    # that can easily cause instability.
+    x = x - x.mean(dim=1, keepdim=True)
+    # x_covar: (num_groups, channels_per_group, channels_per_group)
+    x_covar = torch.matmul(x.transpose(1, 2), x)
+    x_covar_mean_diag = _diag(x_covar).mean()
+    # the following expression is what we'd get if we took the matrix product
+    # of each covariance and measured the mean of its trace, i.e.
+    # the same as _diag(torch.matmul(x_covar, x_covar)).mean().
+    x_covarsq_mean_diag = (x_covar**2).sum() / (num_groups * channels_per_group)
+    # this metric will be >= 1.0; the larger it is, the less 'white' the data was.
+    metric = x_covarsq_mean_diag / (x_covar_mean_diag**2 + 1.0e-20)
+    return metric
+
+
+class WhiteningPenaltyFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, module: nn.Module) -> Tensor:
+        ctx.save_for_backward(x)
+        ctx.module = module
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor):
+        (x_orig,) = ctx.saved_tensors
+        w = ctx.module
+
+        try:
+            with torch.enable_grad():
+                with torch.cuda.amp.autocast(enabled=False):
+                    x_detached = x_orig.to(torch.float32).detach()
+                    x_detached.requires_grad = True
+
+                    metric = _whitening_metric(x_detached, w.num_groups)
+
+                    if random.random() < 0.005 or __name__ == "__main__":
+                        logging.info(
+                            f"Whitening: name={w.name}, num_groups={w.num_groups}, num_channels={x_orig.shape[-1]}, "
+                            f"metric={metric.item():.2f} vs. limit={float(w.whitening_limit)}"
+                        )
+
+                    if metric < float(w.whitening_limit):
+                        w.prob = w.min_prob
+                        return x_grad, None
+                    else:
+                        w.prob = w.max_prob
+                        metric.backward()
+                        penalty_grad = x_detached.grad
+                        scale = w.grad_scale * (
+                            x_grad.to(torch.float32).norm()
+                            / (penalty_grad.norm() + 1.0e-20)
+                        )
+                        penalty_grad = penalty_grad * scale
+                        return x_grad + penalty_grad.to(x_grad.dtype), None
+        except Exception as e:
+            logging.info(
+                f"Caught exception in Whiten backward: {e}, size={list(x_grad.shape)}, will continue."
+            )
+        return x_grad, None
+
+
+class Whiten(nn.Module):
+    def __init__(
+        self,
+        num_groups: int,
+        whitening_limit: FloatLike,
+        prob: Union[float, Tuple[float, float]],
+        grad_scale: FloatLike,
+    ):
+        """
+        Args:
+          num_groups: the number of groups to divide the channel dim into before
+            whitening.  We will attempt to make the feature covariance
+            within each group, after mean subtraction, as "white" as possible,
+            while having the same trace across all groups.
+         whitening_limit: a value greater than 1.0, that dictates how much
+           freedom we have to violate the constraints.  1.0 would mean perfectly
+           white, with exactly the same trace across groups; larger values
+           give more freedom.  E.g. 2.0.
+         prob: the probability with which we apply the gradient modification
+           (also affects the grad scale).  May be supplied as a float,
+           or as a pair (min_prob, max_prob)
+
+          grad_scale: determines the scale on the gradient term from this object,
+            relative to the rest of the gradient on the attention weights.
+            E.g. 0.02 (you may want to use smaller values than this if prob is large)
+        """
+        super(Whiten, self).__init__()
+        assert num_groups >= 1
+        assert float(whitening_limit) >= 1
+        assert grad_scale >= 0
+        self.num_groups = num_groups
+        self.whitening_limit = whitening_limit
+        self.grad_scale = grad_scale
+
+        if isinstance(prob, float):
+            prob = (prob, prob)
+        (self.min_prob, self.max_prob) = prob
+        assert 0 < self.min_prob <= self.max_prob <= 1
+        self.prob = self.max_prob
+        self.name = None  # will be set in training loop
+
+    def forward(self, x: Tensor) -> Tensor:
+        """
+        In the forward pass, this function just returns the input unmodified.
+        In the backward pass, it will modify the gradients to ensure that the
+        distribution in each group has close to (lambda times I) as the covariance
+        after mean subtraction, with the same lambda across groups.
+        For whitening_limit > 1, there will be more freedom to violate this
+        constraint.
+
+        Args:
+           x: the input of shape (*, num_channels)
+
+        Returns:
+            x, unmodified.   You should make sure
+        you use the returned value, or the graph will be freed
+        and nothing will happen in backprop.
+        """
+        grad_scale = float(self.grad_scale)
+        if not x.requires_grad or random.random() > self.prob or grad_scale == 0:
+            return _no_op(x)
+        else:
+            return WhiteningPenaltyFunction.apply(x, self)
+
+
+class WithLoss(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, y: Tensor, name: str):
+        ctx.y_shape = y.shape
+        if random.random() < 0.002 and name is not None:
+            loss_sum = y.sum().item()
+            logging.info(f"WithLoss: name={name}, loss-sum={loss_sum:.3e}")
+        return x
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor):
+        return (
+            ans_grad,
+            torch.ones(ctx.y_shape, dtype=ans_grad.dtype, device=ans_grad.device),
+            None,
+        )
+
+
+def with_loss(x, y, name):
+    # returns x but adds y.sum() to the loss function.
+    return WithLoss.apply(x, y, name)
+
+
+class ScaleGradFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, alpha: float) -> Tensor:
+        ctx.alpha = alpha
+        return x
+
+    @staticmethod
+    def backward(ctx, grad: Tensor):
+        return grad * ctx.alpha, None
+
+
+def scale_grad(x: Tensor, alpha: float):
+    return ScaleGradFunction.apply(x, alpha)
+
+
+class ScaleGrad(nn.Module):
+    def __init__(self, alpha: float):
+        super().__init__()
+        self.alpha = alpha
+
+    def forward(self, x: Tensor) -> Tensor:
+        return scale_grad(x, self.alpha)
+
+
+class LimitParamValue(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, min: float, max: float):
+        ctx.save_for_backward(x)
+        assert max >= min
+        ctx.min = min
+        ctx.max = max
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor):
+        (x,) = ctx.saved_tensors
+        # where x < ctx.min, ensure all grads are negative (this will tend to make
+        # x more positive).
+        x_grad = x_grad * torch.where(
+            torch.logical_and(x_grad > 0, x < ctx.min), -1.0, 1.0
+        )
+        # where x > ctx.max, ensure all grads are positive (this will tend to make
+        # x more negative).
+        x_grad *= torch.where(torch.logical_and(x_grad < 0, x > ctx.max), -1.0, 1.0)
+        return x_grad, None, None
+
+
+def limit_param_value(
+    x: Tensor, min: float, max: float, prob: float = 0.6, training: bool = True
+):
+    # You apply this to (typically) an nn.Parameter during training to ensure that its
+    # (elements mostly) stays within a supplied range.  This is done by modifying the
+    # gradients in backprop.
+    # It's not necessary to do this on every batch: do it only some of the time,
+    # to save a little time.
+    if training and random.random() < prob:
+        return LimitParamValue.apply(x, min, max)
+    else:
+        return x
+
+
+def _no_op(x: Tensor) -> Tensor:
+    if torch.jit.is_scripting():
+        return x
+    else:
+        # a no-op function that will have a node in the autograd graph,
+        # to avoid certain bugs relating to backward hooks
+        return x.chunk(1, dim=-1)[0]
+
+
+class Identity(torch.nn.Module):
+    def __init__(self):
+        super(Identity, self).__init__()
+
+    def forward(self, x):
+        return _no_op(x)
+
+
+class DoubleSwishFunction(torch.autograd.Function):
+    """
+      double_swish(x) = x * torch.sigmoid(x-1)
+
+    This is a definition, originally motivated by its close numerical
+    similarity to swish(swish(x)), where swish(x) =  x * sigmoid(x).
+
+    Memory-efficient derivative computation:
+     double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1)
+     double_swish'(x) = d/dx double_swish(x) =  x * s'(x) + x' * s(x) = x * s'(x) + s(x).
+     Now, s'(x) = s(x) * (1-s(x)).
+     double_swish'(x) =  x * s'(x) + s(x).
+                      =  x * s(x) * (1-s(x)) + s(x).
+                     = double_swish(x) * (1-s(x)) + s(x)
+     ... so we just need to remember s(x) but not x itself.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+        x_dtype = x.dtype
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        s = torch.sigmoid(x - 1.0)
+        y = x * s
+
+        if requires_grad:
+            deriv = y * (1 - s) + s
+
+            # notes on derivative of x * sigmoid(x - 1):
+            # https://www.wolframalpha.com/input?i=d%2Fdx+%28x+*+sigmoid%28x-1%29%29
+            # min \simeq -0.043638.  Take floor as -0.044 so it's a lower bund
+            # max \simeq 1.1990.   Take ceil to be 1.2 so it's an upper bound.
+            # the combination of "+ torch.rand_like(deriv)" and casting to torch.uint8 (which
+            # floors), should be expectation-preserving.
+            floor = -0.044
+            ceil = 1.2
+            d_scaled = (deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                deriv
+            )
+            if __name__ == "__main__":
+                # for self-testing only.
+                assert d_scaled.min() >= 0.0
+                assert d_scaled.max() < 256.0
+            d_int = d_scaled.to(torch.uint8)
+            ctx.save_for_backward(d_int)
+        if x.dtype == torch.float16 or torch.is_autocast_enabled():
+            y = y.to(torch.float16)
+        return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+        floor = -0.043637
+        ceil = 1.2
+
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class DoubleSwish(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Return double-swish activation function which is an approximation to Swish(Swish(x)),
+        that we approximate closely with x * sigmoid(x-1).
+        """
+        if torch.jit.is_scripting():
+            return x * torch.sigmoid(x - 1.0)
+        return DoubleSwishFunction.apply(x)
+
+
+# Dropout2 is just like normal dropout, except it supports schedules on the dropout rates.
+class Dropout2(nn.Module):
+    def __init__(self, p: FloatLike):
+        super().__init__()
+        self.p = p
+
+    def forward(self, x: Tensor) -> Tensor:
+        return torch.nn.functional.dropout(x, p=float(self.p), training=self.training)
+
+
+class MulForDropout3(torch.autograd.Function):
+    # returns (x * y * alpha) where alpha is a float and y doesn't require
+    # grad and is zero-or-one.
+    @staticmethod
+    @custom_fwd
+    def forward(ctx, x, y, alpha):
+        assert not y.requires_grad
+        ans = x * y * alpha
+        ctx.save_for_backward(ans)
+        ctx.alpha = alpha
+        return ans
+
+    @staticmethod
+    @custom_bwd
+    def backward(ctx, ans_grad):
+        (ans,) = ctx.saved_tensors
+        x_grad = ctx.alpha * ans_grad * (ans != 0)
+        return x_grad, None, None
+
+
+# Dropout3 is just like normal dropout, except it supports schedules on the dropout rates,
+# and it lets you choose one dimension to share the dropout mask over
+class Dropout3(nn.Module):
+    def __init__(self, p: FloatLike, shared_dim: int):
+        super().__init__()
+        self.p = p
+        self.shared_dim = shared_dim
+
+    def forward(self, x: Tensor) -> Tensor:
+        p = float(self.p)
+        if not self.training or p == 0:
+            return _no_op(x)
+        scale = 1.0 / (1 - p)
+        rand_shape = list(x.shape)
+        rand_shape[self.shared_dim] = 1
+        mask = torch.rand(*rand_shape, device=x.device) > p
+        ans = MulForDropout3.apply(x, mask, scale)
+        return ans
+
+
+class SwooshLFunction(torch.autograd.Function):
+    """
+    swoosh(x) =  log(1 + exp(x-4)) - 0.08*x - 0.035
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+        x_dtype = x.dtype
+
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+
+        coeff = -0.08
+
+        with torch.cuda.amp.autocast(enabled=False):
+            with torch.enable_grad():
+                x = x.detach()
+                x.requires_grad = True
+                y = torch.logaddexp(zero, x - 4.0) + coeff * x - 0.035
+
+                if not requires_grad:
+                    return y
+                y.backward(gradient=torch.ones_like(y))
+
+                grad = x.grad
+                floor = coeff
+                ceil = 1.0 + coeff + 0.005
+
+                d_scaled = (grad - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                    grad
+                )
+                if __name__ == "__main__":
+                    # for self-testing only.
+                    assert d_scaled.min() >= 0.0
+                    assert d_scaled.max() < 256.0
+
+                d_int = d_scaled.to(torch.uint8)
+                ctx.save_for_backward(d_int)
+                if x.dtype == torch.float16 or torch.is_autocast_enabled():
+                    y = y.to(torch.float16)
+                return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+
+        coeff = -0.08
+        floor = coeff
+        ceil = 1.0 + coeff + 0.005
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class SwooshL(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swoosh-L activation."""
+        if torch.jit.is_scripting():
+            zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+            return torch.logaddexp(zero, x - 4.0) - 0.08 * x - 0.035
+        if not x.requires_grad:
+            return k2.swoosh_l_forward(x)
+        else:
+            return k2.swoosh_l(x)
+        # return SwooshLFunction.apply(x)
+
+
+class SwooshRFunction(torch.autograd.Function):
+    """
+     swoosh(x) =  log(1 + exp(x-1)) - 0.08*x - 0.313261687
+
+    derivatives are between -0.08 and 0.92.
+
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+        x_dtype = x.dtype
+
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            with torch.enable_grad():
+                x = x.detach()
+                x.requires_grad = True
+                y = torch.logaddexp(zero, x - 1.0) - 0.08 * x - 0.313261687
+
+                if not requires_grad:
+                    return y
+                y.backward(gradient=torch.ones_like(y))
+
+                grad = x.grad
+                floor = -0.08
+                ceil = 0.925
+
+                d_scaled = (grad - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                    grad
+                )
+                if __name__ == "__main__":
+                    # for self-testing only.
+                    assert d_scaled.min() >= 0.0
+                    assert d_scaled.max() < 256.0
+                d_int = d_scaled.to(torch.uint8)
+                ctx.save_for_backward(d_int)
+                if x.dtype == torch.float16 or torch.is_autocast_enabled():
+                    y = y.to(torch.float16)
+                return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+        floor = -0.08
+        ceil = 0.925
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class SwooshR(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return Swoosh-R activation."""
+        if torch.jit.is_scripting():
+            zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
+            return torch.logaddexp(zero, x - 1.0) - 0.08 * x - 0.313261687
+        if not x.requires_grad:
+            return k2.swoosh_r_forward(x)
+        else:
+            return k2.swoosh_r(x)
+        # return SwooshRFunction.apply(x)
+
+
+# simple version of SwooshL that does not redefine the backprop, used in
+# ActivationDropoutAndLinearFunction.
+def SwooshLForward(x: Tensor):
+    x_offset = x - 4.0
+    log_sum = (1.0 + x_offset.exp()).log().to(x.dtype)
+    log_sum = torch.where(log_sum == float("inf"), x_offset, log_sum)
+    return log_sum - 0.08 * x - 0.035
+
+
+# simple version of SwooshR that does not redefine the backprop, used in
+# ActivationDropoutAndLinearFunction.
+def SwooshRForward(x: Tensor):
+    x_offset = x - 1.0
+    log_sum = (1.0 + x_offset.exp()).log().to(x.dtype)
+    log_sum = torch.where(log_sum == float("inf"), x_offset, log_sum)
+    return log_sum - 0.08 * x - 0.313261687
+
+
+class ActivationDropoutAndLinearFunction(torch.autograd.Function):
+    @staticmethod
+    @custom_fwd
+    def forward(
+        ctx,
+        x: Tensor,
+        weight: Tensor,
+        bias: Optional[Tensor],
+        activation: str,
+        dropout_p: float,
+        dropout_shared_dim: Optional[int],
+    ):
+        if dropout_p != 0.0:
+            dropout_shape = list(x.shape)
+            if dropout_shared_dim is not None:
+                dropout_shape[dropout_shared_dim] = 1
+            # else it won't be very memory efficient.
+            dropout_mask = (1.0 / (1.0 - dropout_p)) * (
+                torch.rand(*dropout_shape, device=x.device, dtype=x.dtype) > dropout_p
+            )
+        else:
+            dropout_mask = None
+
+        ctx.save_for_backward(x, weight, bias, dropout_mask)
+
+        ctx.activation = activation
+
+        forward_activation_dict = {
+            "SwooshL": k2.swoosh_l_forward,
+            "SwooshR": k2.swoosh_r_forward,
+        }
+        # it will raise a KeyError if this fails.  This will be an error.  We let it
+        # propagate to the user.
+        activation_func = forward_activation_dict[activation]
+        x = activation_func(x)
+        if dropout_mask is not None:
+            x = x * dropout_mask
+        x = torch.nn.functional.linear(x, weight, bias)
+        return x
+
+    @staticmethod
+    @custom_bwd
+    def backward(ctx, ans_grad: Tensor):
+        saved = ctx.saved_tensors
+        (x, weight, bias, dropout_mask) = saved
+
+        forward_and_deriv_activation_dict = {
+            "SwooshL": k2.swoosh_l_forward_and_deriv,
+            "SwooshR": k2.swoosh_r_forward_and_deriv,
+        }
+        # the following lines a KeyError if the activation is unrecognized.
+        # This will be an error.  We let it propagate to the user.
+        func = forward_and_deriv_activation_dict[ctx.activation]
+
+        y, func_deriv = func(x)
+        if dropout_mask is not None:
+            y = y * dropout_mask
+        # now compute derivative of y w.r.t. weight and bias..
+        # y: (..., in_channels), ans_grad: (..., out_channels),
+        (out_channels, in_channels) = weight.shape
+
+        in_channels = y.shape[-1]
+        g = ans_grad.reshape(-1, out_channels)
+        weight_deriv = torch.matmul(g.t(), y.reshape(-1, in_channels))
+        y_deriv = torch.matmul(ans_grad, weight)
+        bias_deriv = None if bias is None else g.sum(dim=0)
+        x_deriv = y_deriv * func_deriv
+        if dropout_mask is not None:
+            # order versus func_deriv does not matter
+            x_deriv = x_deriv * dropout_mask
+
+        return x_deriv, weight_deriv, bias_deriv, None, None, None
+
+
+class ActivationDropoutAndLinear(torch.nn.Module):
+    """
+     This merges an activation function followed by dropout and then a nn.Linear module;
+     it does so in a memory efficient way so that it only stores the input to the whole
+     module.  If activation == SwooshL and dropout_shared_dim != None, this will be
+     equivalent to:
+       nn.Sequential(SwooshL(),
+                     Dropout3(dropout_p, shared_dim=dropout_shared_dim),
+                     ScaledLinear(in_channels, out_channels, bias=bias,
+                                  initial_scale=initial_scale))
+    If dropout_shared_dim is None, the dropout would be equivalent to
+    Dropout2(dropout_p).  Note: Dropout3 will be more memory efficient as the dropout
+    mask is smaller.
+
+     Args:
+        in_channels: number of input channels, e.g. 256
+        out_channels: number of output channels, e.g. 256
+        bias: if true, have a bias
+        activation: the activation function, for now just support SwooshL.
+        dropout_p: the dropout probability or schedule (happens after nonlinearity).
+        dropout_shared_dim: the dimension, if any, across which the dropout mask is
+             shared (e.g. the time dimension).  If None, this may be less memory
+             efficient if there are modules before this one that cache the input
+             for their backprop (e.g. Balancer or Whiten).
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        bias: bool = True,
+        activation: str = "SwooshL",
+        dropout_p: FloatLike = 0.0,
+        dropout_shared_dim: Optional[int] = -1,
+        initial_scale: float = 1.0,
+    ):
+        super().__init__()
+        # create a temporary module of nn.Linear that we'll steal the
+        # weights and bias from
+        layer = ScaledLinear(
+            in_channels, out_channels, bias=bias, initial_scale=initial_scale
+        )
+
+        self.weight = layer.weight
+        # register_parameter properly handles making it a parameter when l.bias
+        # is None. I think there is some reason for doing it this way rather
+        # than just setting it to None but I don't know what it is, maybe
+        # something to do with exporting the module..
+        self.register_parameter("bias", layer.bias)
+
+        self.activation = activation
+        self.dropout_p = dropout_p
+        self.dropout_shared_dim = dropout_shared_dim
+
+    def forward(self, x: Tensor):
+        if torch.jit.is_scripting():
+            if self.activation == "SwooshL":
+                x = SwooshLForward(x)
+            elif self.activation == "SwooshR":
+                x = SwooshRForward(x)
+            else:
+                assert False, self.activation
+            return torch.nn.functional.linear(x, self.weight, self.bias)
+
+        return ActivationDropoutAndLinearFunction.apply(
+            x,
+            self.weight,
+            self.bias,
+            self.activation,
+            float(self.dropout_p),
+            self.dropout_shared_dim,
+        )
+
+
+class ClipGradFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, limit: float):
+        ctx.limit = limit
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad, *args):
+        return x_grad.clamp(-ctx.limit, ctx.limit), None
+
+
+def clip_grad(x: Tensor, limit: float):
+    return ClipGradFunction.apply(x, limit)
+
+
+class AbsValuePenalizer(nn.Module):
+    """
+    This module adds a penalty to the loss function when ever the absolute value of
+    any element of the input tensor exceeds a certain limit.
+    """
+
+    def __init__(self, limit: float, prob: float = 0.1, penalty: float = 1.0e-04):
+        super().__init__()
+        self.limit = limit
+        self.penalty = penalty
+
+        self.prob = prob
+        self.name = None  # will be set in training loop
+
+        # 20% of the time we will return and do nothing because memory usage is
+        # too high.
+        self.mem_cutoff = CutoffEstimator(0.2)
+
+    def forward(self, x: Tensor) -> Tensor:
+        if (
+            torch.jit.is_scripting()
+            or not x.requires_grad
+            or not self.training
+            or random.random() > self.prob
+        ):
+            # or (x.is_cuda and self.mem_cutoff(torch.cuda.memory_allocated()))
+            return _no_op(x)  # the _no_op op is to make our diagnostics code work.
+
+        x = penalize_abs_values_gt(
+            x, limit=self.limit, penalty=self.penalty, name=self.name
+        )
+        return x
+
+
+def convert_num_channels(x: Tensor, num_channels: int) -> Tensor:
+    if num_channels <= x.shape[-1]:
+        return x[..., :num_channels]
+    else:
+        shape = list(x.shape)
+        shape[-1] = num_channels - shape[-1]
+        zeros = torch.zeros(*shape, dtype=x.dtype, device=x.device)
+        return torch.cat((x, zeros), dim=-1)
+
+
+def _test_whiten():
+    for proportion in [0.1, 0.5, 10.0]:
+        logging.info(f"_test_whiten(): proportion = {proportion}")
+        x = torch.randn(100, 128)
+        direction = torch.randn(128)
+        coeffs = torch.randn(100, 1)
+        x += proportion * direction * coeffs
+
+        x.requires_grad = True
+
+        num_channels = 128
+        m = Whiten(
+            1, 5.0, prob=1.0, grad_scale=0.1  # num_groups  # whitening_limit,
+        )  # grad_scale
+
+        for _ in range(4):
+            y = m(x)
+
+        y_grad = torch.randn_like(x)
+        y.backward(gradient=y_grad)
+
+        if proportion < 0.2:
+            assert torch.allclose(x.grad, y_grad)
+        elif proportion > 1.0:
+            assert not torch.allclose(x.grad, y_grad)
+
+
+def _test_balancer_sign():
+    probs = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = 1.0 * ((2.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))) - 1.0)
+    x = x.detach()
+    x.requires_grad = True
+    m = Balancer(
+        probs.numel(),
+        channel_dim=0,
+        min_positive=0.05,
+        max_positive=0.95,
+        min_abs=0.0,
+        prob=1.0,
+    )
+
+    y_grad = torch.sign(torch.randn(probs.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_balancer_sign: x = ", x)
+    print("_test_balancer_sign: y grad = ", y_grad)
+    print("_test_balancer_sign: x grad = ", x.grad)
+
+
+def _test_balancer_magnitude():
+    magnitudes = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(-1)
+    x = x.detach()
+    x.requires_grad = True
+    m = Balancer(
+        magnitudes.numel(),
+        channel_dim=0,
+        min_positive=0.0,
+        max_positive=1.0,
+        min_abs=0.2,
+        max_abs=0.7,
+        prob=1.0,
+    )
+
+    y_grad = torch.sign(torch.randn(magnitudes.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_balancer_magnitude: x = ", x)
+    print("_test_balancer_magnitude: y grad = ", y_grad)
+    print("_test_balancer_magnitude: x grad = ", x.grad)
+
+
+def _test_double_swish_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    m = DoubleSwish()
+
+    tol = (1.2 - (-0.043637)) / 255.0
+    torch.autograd.gradcheck(m, x, atol=tol)
+
+    # for self-test.
+    x = torch.randn(1000, 1000, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    y = m(x)
+
+
+def _test_swooshl_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    m = SwooshL()
+
+    tol = 1.0 / 255.0
+    torch.autograd.gradcheck(m, x, atol=tol, eps=0.01)
+
+    # for self-test.
+    x = torch.randn(1000, 1000, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    y = m(x)
+
+
+def _test_swooshr_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    m = SwooshR()
+
+    tol = 1.0 / 255.0
+    torch.autograd.gradcheck(m, x, atol=tol, eps=0.01)
+
+    # for self-test.
+    x = torch.randn(1000, 1000, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    y = m(x)
+
+
+def _test_softmax():
+    a = torch.randn(2, 10, dtype=torch.float64)
+    b = a.clone()
+    a.requires_grad = True
+    b.requires_grad = True
+    a.softmax(dim=1)[:, 0].sum().backward()
+    print("a grad = ", a.grad)
+    softmax(b, dim=1)[:, 0].sum().backward()
+    print("b grad = ", b.grad)
+    assert torch.allclose(a.grad, b.grad)
+
+
+def _test_piecewise_linear():
+    p = PiecewiseLinear((0, 10.0))
+    for x in [-100, 0, 100]:
+        assert p(x) == 10.0
+    p = PiecewiseLinear((0, 10.0), (1, 0.0))
+    for x, y in [(-100, 10.0), (0, 10.0), (0.5, 5.0), (1, 0.0), (2, 0.0)]:
+        print("x, y = ", x, y)
+        assert p(x) == y, (x, p(x), y)
+
+    q = PiecewiseLinear((0.5, 15.0), (0.6, 1.0))
+    x_vals = [-1.0, 0.0, 0.1, 0.2, 0.5, 0.6, 0.7, 0.9, 1.0, 2.0]
+    pq = p.max(q)
+    for x in x_vals:
+        y1 = max(p(x), q(x))
+        y2 = pq(x)
+        assert abs(y1 - y2) < 0.001
+    pq = p.min(q)
+    for x in x_vals:
+        y1 = min(p(x), q(x))
+        y2 = pq(x)
+        assert abs(y1 - y2) < 0.001
+    pq = p + q
+    for x in x_vals:
+        y1 = p(x) + q(x)
+        y2 = pq(x)
+        assert abs(y1 - y2) < 0.001
+
+
+def _test_activation_dropout_and_linear():
+    in_channels = 20
+    out_channels = 30
+
+    for bias in [True, False]:
+        # actually we don't test for dropout_p != 0.0 because forward functions will give
+        # different answers.  This is because we are using the k2 implementation of
+        # swoosh_l an swoosh_r inside SwooshL() and SwooshR(), and they call randn()
+        # internally, messing up the random state.
+        for dropout_p in [0.0]:
+            for activation in ["SwooshL", "SwooshR"]:
+                m1 = nn.Sequential(
+                    SwooshL() if activation == "SwooshL" else SwooshR(),
+                    Dropout3(p=dropout_p, shared_dim=-1),
+                    ScaledLinear(
+                        in_channels, out_channels, bias=bias, initial_scale=0.5
+                    ),
+                )
+                m2 = ActivationDropoutAndLinear(
+                    in_channels,
+                    out_channels,
+                    bias=bias,
+                    initial_scale=0.5,
+                    activation=activation,
+                    dropout_p=dropout_p,
+                )
+                with torch.no_grad():
+                    m2.weight[:] = m1[2].weight
+                    if bias:
+                        m2.bias[:] = m1[2].bias
+                # make sure forward gives same result.
+                x1 = torch.randn(10, in_channels)
+                x1.requires_grad = True
+
+                # TEMP.
+                assert torch.allclose(
+                    SwooshRFunction.apply(x1), SwooshRForward(x1), atol=1.0e-03
+                )
+
+                x2 = x1.clone().detach()
+                x2.requires_grad = True
+                seed = 10
+                torch.manual_seed(seed)
+                y1 = m1(x1)
+                y_grad = torch.randn_like(y1)
+                y1.backward(gradient=y_grad)
+                torch.manual_seed(seed)
+                y2 = m2(x2)
+                y2.backward(gradient=y_grad)
+
+                print(
+                    f"bias = {bias}, dropout_p = {dropout_p}, activation = {activation}"
+                )
+                print("y1 = ", y1)
+                print("y2 = ", y2)
+                assert torch.allclose(y1, y2, atol=0.02)
+                assert torch.allclose(m1[2].weight.grad, m2.weight.grad, atol=1.0e-05)
+                if bias:
+                    assert torch.allclose(m1[2].bias.grad, m2.bias.grad, atol=1.0e-05)
+                print("x1.grad = ", x1.grad)
+                print("x2.grad = ", x2.grad)
+
+                def isclose(a, b):
+                    # return true if cosine similarity is > 0.9.
+                    return (a * b).sum() > 0.9 * (
+                        (a**2).sum() * (b**2).sum()
+                    ).sqrt()
+
+                # the SwooshL() implementation has a noisy gradient due to 1-byte
+                # storage of it.
+                assert isclose(x1.grad, x2.grad)
+
+
+if __name__ == "__main__":
+    logging.getLogger().setLevel(logging.INFO)
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    _test_piecewise_linear()
+    _test_softmax()
+    _test_whiten()
+    _test_balancer_sign()
+    _test_balancer_magnitude()
+    _test_swooshl_deriv()
+    _test_swooshr_deriv()
+    _test_activation_dropout_and_linear()
+    _test_double_swish_deriv()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/subsampling.py b/egs/libriheavy/ASR/zipformer_prompt_asr/subsampling.py
new file mode 100644
index 000000000..7acbc1808
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/subsampling.py
@@ -0,0 +1,276 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from typing import Tuple
+
+import torch
+from scaling import (
+    Balancer,
+    BiasNorm,
+    Dropout3,
+    FloatLike,
+    Optional,
+    ScaledConv2d,
+    ScaleGrad,
+    ScheduledFloat,
+    SwooshL,
+    SwooshR,
+    Whiten,
+)
+from torch import Tensor, nn
+
+
+class ConvNeXt(nn.Module):
+    """
+    Our interpretation of the ConvNeXt module as used in https://arxiv.org/pdf/2206.14747.pdf
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        hidden_ratio: int = 3,
+        kernel_size: Tuple[int, int] = (7, 7),
+        layerdrop_rate: FloatLike = None,
+    ):
+        super().__init__()
+        padding = ((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2)
+        hidden_channels = channels * hidden_ratio
+        if layerdrop_rate is None:
+            layerdrop_rate = ScheduledFloat((0.0, 0.2), (20000.0, 0.015))
+        self.layerdrop_rate = layerdrop_rate
+
+        self.depthwise_conv = nn.Conv2d(
+            in_channels=channels,
+            out_channels=channels,
+            groups=channels,
+            kernel_size=kernel_size,
+            padding=padding,
+        )
+
+        self.pointwise_conv1 = nn.Conv2d(
+            in_channels=channels, out_channels=hidden_channels, kernel_size=1
+        )
+
+        self.hidden_balancer = Balancer(
+            hidden_channels,
+            channel_dim=1,
+            min_positive=0.3,
+            max_positive=1.0,
+            min_abs=0.75,
+            max_abs=5.0,
+        )
+
+        self.activation = SwooshL()
+        self.pointwise_conv2 = ScaledConv2d(
+            in_channels=hidden_channels,
+            out_channels=channels,
+            kernel_size=1,
+            initial_scale=0.01,
+        )
+
+        self.out_balancer = Balancer(
+            channels,
+            channel_dim=1,
+            min_positive=0.4,
+            max_positive=0.6,
+            min_abs=1.0,
+            max_abs=6.0,
+        )
+        self.out_whiten = Whiten(
+            num_groups=1,
+            whitening_limit=5.0,
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        if torch.jit.is_scripting() or not self.training:
+            return self.forward_internal(x)
+        layerdrop_rate = float(self.layerdrop_rate)
+
+        if layerdrop_rate != 0.0:
+            batch_size = x.shape[0]
+            mask = (
+                torch.rand((batch_size, 1, 1, 1), dtype=x.dtype, device=x.device)
+                > layerdrop_rate
+            )
+        else:
+            mask = None
+        # turns out this caching idea does not work with --world-size > 1
+        # return caching_eval(self.forward_internal, x, mask)
+        return self.forward_internal(x, mask)
+
+    def forward_internal(
+        self, x: Tensor, layer_skip_mask: Optional[Tensor] = None
+    ) -> Tensor:
+        """
+        x layout: (N, C, H, W), i.e. (batch_size, num_channels, num_frames, num_freqs)
+
+        The returned value has the same shape as x.
+        """
+        bypass = x
+        x = self.depthwise_conv(x)
+        x = self.pointwise_conv1(x)
+        x = self.hidden_balancer(x)
+        x = self.activation(x)
+        x = self.pointwise_conv2(x)
+
+        if layer_skip_mask is not None:
+            x = x * layer_skip_mask
+
+        x = bypass + x
+        x = self.out_balancer(x)
+        x = x.transpose(1, 3)  # (N, W, H, C); need channel dim to be last
+        x = self.out_whiten(x)
+        x = x.transpose(1, 3)  # (N, C, H, W)
+
+        return x
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/2 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = (T-3)//2 - 2 == (T-7)//2
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+        dropout: FloatLike = 0.1,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, (T-3)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+          bottleneck:
+            bottleneck dimension for 1d squeeze-excite
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        # The ScaleGrad module is there to prevent the gradients
+        # w.r.t. the weight or bias of the first Conv2d module in self.conv from
+        # exceeding the range of fp16 when using automatic mixed precision (amp)
+        # training.  (The second one is necessary to stop its bias from getting
+        # a too-large gradient).
+
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=(0, 1),  # (time, freq)
+            ),
+            ScaleGrad(0.2),
+            Balancer(layer1_channels, channel_dim=1, max_abs=1.0),
+            SwooshR(),
+            nn.Conv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+                padding=0,
+            ),
+            Balancer(layer2_channels, channel_dim=1, max_abs=4.0),
+            SwooshR(),
+            nn.Conv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=(1, 2),  # (time, freq)
+            ),
+            Balancer(layer3_channels, channel_dim=1, max_abs=4.0),
+            SwooshR(),
+        )
+
+        # just one convnext layer
+        self.convnext = ConvNeXt(layer3_channels, kernel_size=(7, 7))
+
+        out_width = (((in_channels - 1) // 2) - 1) // 2
+
+        self.out = nn.Linear(out_width * layer3_channels, out_channels)
+        # use a larger than normal grad_scale on this whitening module; there is
+        # only one such module, so there is not a concern about adding together
+        # many copies of this extra gradient term.
+        self.out_whiten = Whiten(
+            num_groups=1,
+            whitening_limit=ScheduledFloat((0.0, 4.0), (20000.0, 8.0), default=4.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.02,
+        )
+
+        # max_log_eps=0.0 is to prevent both eps and the output of self.out from
+        # getting large, there is an unnecessary degree of freedom.
+        self.out_norm = BiasNorm(out_channels)
+        self.dropout = Dropout3(dropout, shared_dim=1)
+
+    def forward(self, x: torch.Tensor, x_lens: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+
+        Returns:
+          - a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+          - output lengths, of shape (batch_size,)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        # scaling x by 0.1 allows us to use a larger grad-scale in fp16 "amp" (automatic mixed precision)
+        # training, since the weights in the first convolution are otherwise the limiting factor for getting infinite
+        # gradients.
+        x = self.conv(x)
+        x = self.convnext(x)
+
+        # Now x is of shape (N, odim, ((T-3)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+
+        x = x.transpose(1, 2).reshape(b, t, c * f)
+        # now x: (N, ((T-1)//2 - 1))//2, out_width * layer3_channels))
+
+        x = self.out(x)
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_whiten(x)
+        x = self.out_norm(x)
+        x = self.dropout(x)
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            x_lens = (x_lens - 7) // 2
+        assert x.size(1) == x_lens.max().item()
+
+        return x, x_lens
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/test_model.py b/egs/libriheavy/ASR/zipformer_prompt_asr/test_model.py
new file mode 100755
index 000000000..13483637d
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/test_model.py
@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless4/test_model.py
+"""
+
+from scaling import ScheduledFloat
+from train_subformer import get_params, get_text_encoder, get_transducer_model
+from zipformer import Zipformer2
+
+
+def test_model_1():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = 24
+    params.dim_feedforward = 1536  # 384 * 4
+    params.encoder_dim = 384
+    model = get_transducer_model(params)
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+# See Table 1 from https://arxiv.org/pdf/2005.08100.pdf
+def test_model_M():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.num_encoder_layers = "2,4,3,2,4"
+    params.feedforward_dims = "1024,1024,2048,2048,1024"
+    params.nhead = "8,8,8,8,8"
+    params.encoder_dims = "384,384,384,384,384"
+    params.attention_dims = "192,192,192,192,192"
+    params.encoder_unmasked_dims = "256,256,256,256,256"
+    params.zipformer_downsampling_factors = "1,2,4,8,2"
+    params.cnn_module_kernels = "31,31,15,15"
+
+    params.text_encoder_dim = (192, 192, 256, 384)
+    params.decoder_dim = 512
+    params.joiner_dim = 512
+    model = Zipformer2(
+        output_downsampling_factor=8,
+        downsampling_factor=(1, 2, 4, 8),
+        num_encoder_layers=(2, 4, 4, 4),
+        encoder_dim=(192, 192, 256, 384),
+        encoder_unmasked_dim=(192, 192, 256, 256),
+        query_head_dim=(32, 32, 32, 32),
+        pos_head_dim=(4, 4, 4, 4),
+        value_head_dim=(12, 12, 12, 12),
+        pos_dim=48,
+        num_heads=(4, 4, 4, 8),
+        feedforward_dim=(
+            384,
+            512,
+            768,
+            1024,
+        ),  # could increase this if there is nough data
+        cnn_module_kernel=(31, 31, 15, 15),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=False,
+    )
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+    model = Zipformer2(
+        output_downsampling_factor=8,
+        downsampling_factor=(1, 2, 4, 8),
+        num_encoder_layers=(2, 4, 6, 6),
+        encoder_dim=(256, 256, 384, 512),
+        encoder_unmasked_dim=(196, 196, 256, 256),
+        query_head_dim=(32, 32, 32, 32),
+        pos_head_dim=(4, 4, 4, 4),
+        value_head_dim=(12, 12, 12, 12),
+        pos_dim=48,
+        num_heads=(4, 4, 4, 8),
+        feedforward_dim=(
+            384,
+            512,
+            768,
+            1024,
+        ),  # could increase this if there is nough data
+        cnn_module_kernel=(31, 31, 15, 15),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=False,
+    )
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+
+
+def main():
+    #  test_model_1()
+    test_model_M()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/text_normalization.py b/egs/libriheavy/ASR/zipformer_prompt_asr/text_normalization.py
new file mode 100644
index 000000000..efb4acc3c
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/text_normalization.py
@@ -0,0 +1,101 @@
+#      Copyright      2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+from typing import List
+
+
+def train_text_normalization(s: str) -> str:
+    # replace full-width with half-width
+    s = s.replace("“", '"')
+    s = s.replace("”", '"')
+    s = s.replace("‘", "'")
+    s = s.replace("’", "'")
+    if s[:2] == '" ':  # remove the starting double quote
+        s = s[2:]
+
+    return s
+
+
+def ref_text_normalization(ref_text: str) -> str:
+    # Rule 1: Remove the [FN#[]]
+    p = r"[FN#[0-9]*]"
+    pattern = re.compile(p)
+
+    res = pattern.findall(ref_text)
+    ref_text = re.sub(p, "", ref_text)
+
+    ref_text = train_text_normalization(ref_text)
+
+    return ref_text
+
+
+def remove_non_alphabetic(text: str, strict: bool = True) -> str:
+    # Recommend to set strict to False
+    if not strict:
+        # Note, this also keeps space, single quote(') and hypen (-)
+        text = text.replace("-", " ")
+        text = text.replace("—", " ")
+        return re.sub(r"[^a-zA-Z0-9\s']+", "", text)
+    else:
+        # only keeps space
+        return re.sub(r"[^a-zA-Z\s]+", "", text)
+
+
+def upper_only_alpha(text: str) -> str:
+    return remove_non_alphabetic(text.upper(), strict=False)
+
+
+def lower_only_alpha(text: str) -> str:
+    return remove_non_alphabetic(text.lower(), strict=False)
+
+
+def lower_all_char(text: str) -> str:
+    return text.lower()
+
+
+def upper_all_char(text: str) -> str:
+    return text.upper()
+
+
+def _apply_style_transform(text: List[str], transform: str) -> List[str]:
+    """Apply transform to a list of text. By default, the text are in
+    ground truth format, i.e mixed-punc.
+
+    Args:
+        text (List[str]): Input text string
+        transform (str): Transform to be applied
+
+    Returns:
+        List[str]: _description_
+    """
+    if transform == "mixed-punc":
+        return text
+    elif transform == "upper-no-punc":
+        return [upper_only_alpha(s) for s in text]
+    elif transform == "lower-no-punc":
+        return [lower_only_alpha(s) for s in text]
+    elif transform == "lower-punc":
+        return [lower_all_char(s) for s in text]
+    else:
+        raise NotImplementedError(f"Unseen transform: {transform}")
+
+
+if __name__ == "__main__":
+    ref_text = "Mixed-case English transcription, with punctuation. Actually, it is fully not related."
+    print(ref_text)
+    res = upper_only_alpha(ref_text)
+    print(res)
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py b/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py
new file mode 100644
index 000000000..32302602c
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py
@@ -0,0 +1,1419 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo
+#                                                       Zengwei Yao,
+#                                                       Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+
+# For mix precision training, using MCP style transcript:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./zipformer_prompt_asr/train_baseline.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer_prompt_asr/exp \
+  --transcript-style MCP \
+  --max-duration 1000
+
+# For mix precision training, using UC style transcript:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./zipformer_prompt_asr/train_baseline.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer_prompt_asr/exp \
+  --transcript-style UC \
+  --max-duration 1000
+
+# To train a streaming model
+
+./zipformer_prompt_asr/train_baseline.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --causal 1
+  --exp-dir zipformer/exp \
+  --max-duration 1000
+
+"""
+
+
+import argparse
+import copy
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import k2
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriHeavyAsrDataModule
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model_baseline import Transducer
+from optim import Eden, ScaledAdam
+from scaling import ScheduledFloat
+from subsampling import Conv2dSubsampling
+from text_normalization import train_text_normalization, upper_only_alpha
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_mixed_cased_with_punc(
+    texts: List[str],
+    pre_texts: List[str],
+    context_list: Optional[str] = None,
+    rare_word_list: Optional[List[str]] = None,
+) -> str:
+    # Always get the first one, which is the mixed-cased text with punc
+    out = {"text": texts[0], "pre_text": pre_texts[0]}
+    return out
+
+
+def get_upper_only_alpha(
+    texts: List[str],
+    pre_texts: List[str],
+    context_list: Optional[str] = None,
+    rare_word_list: Optional[List[str]] = None,
+) -> str:
+    # Always get the first one, which is the mixed-cased text with punc,
+    # but with upper case it and remove punctuation
+    out = {
+        "text": upper_only_alpha(texts[0]),
+        "pre_text": upper_only_alpha(pre_texts[0]),
+    }
+    return out
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="Embedding dimension in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--text-encoder-dim",
+        type=str,
+        default="256,256,384,512",
+        help="Embedding dimension in text encoder stacks: a comma-separated list of 4 elements, "
+        "or you should change other configs in the code.",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="Value dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "A single int or comma-separated list.  Must be <= each corresponding encoder_dim.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="Sizes of convolutional kernels in convolution modules in each encoder stack: "
+        "a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. "
+        " Must be just -1 if --causal=False",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="Maximum left-contexts for causal training, measured in frames which will "
+        "be converted to a number of chunks.  If splitting into chunks, "
+        "chunk left-context frames will be chosen randomly from this list; else not relevant.",
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.045, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="Reference batch duration for purposes of adjusting batch counts for setting various "
+        "schedules inside the model",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--transcript-style",
+        type=str,
+        default="UC",
+        choices=["UC", "MCP"],
+        help="""The transcript style used for training. UC stands for upper-cased text w/o punctuations,
+        MCP stands for mix-cased text with punctuation.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(max(params.encoder_dim.split(","))),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    texts = [train_text_normalization(t) for t in texts]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    if random.random() < 0.02:
+        logging.info(f"Ref texts: {texts[0]}")
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            2**22
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    train_cuts = libriheavy.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 30.0:
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].texts[0], out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].texts[0]}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    if params.transcript_style == "UC":
+        text_sampling_func = get_upper_only_alpha
+    else:
+        text_sampling_func = get_mixed_cased_with_punc
+    logging.info(f"Using {params.transcript_style} style for training.")
+    logging.info(f"Text sampling func: {text_sampling_func}")
+    train_dl = libriheavy.train_dataloaders(
+        train_cuts,
+        sampler_state_dict=sampler_state_dict,
+        text_sampling_func=text_sampling_func,
+    )
+
+    valid_cuts = libriheavy.dev_cuts()
+    valid_dl = libriheavy.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py b/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py
new file mode 100755
index 000000000..e253d1118
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py
@@ -0,0 +1,1798 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Xiaoyu Yang,
+#
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+# For mix precision training:
+
+(1) Non-streaming model, **without** context list
+
+./zipformer_prompt_asr/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --subset medium \
+  --causal False \
+  --exp-dir zipformer_prompt_asr/exp \
+  --max-duration 1000 \
+  --memory-layer 0 \
+  --text-encoder-type BERT \
+  --text-encoder-dim 768 \
+  --use-style-prompt True \
+  --use-context-list False
+
+(2) Non-streaming model, **with** context list
+
+./zipformer_prompt_asr/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --subset medium \
+  --causal False \
+  --exp-dir zipformer_prompt_asr/exp \
+  --max-duration 1000 \
+  --memory-layer 0 \
+  --text-encoder-type BERT \
+  --text-encoder-dim 768 \
+  --use-style-prompt True \
+  --use-context-list True \
+  --top-k 10000 \
+  --rare-word-file data/context_biasing/small_rare_words_topk_10000.txt
+
+
+"""
+
+
+import argparse
+import copy
+import logging
+import os
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import k2
+import numpy
+import optim
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriHeavyAsrDataModule
+from dataset import (
+    naive_triplet_text_sampling,
+    random_shuffle_subset,
+    triplet_text_sampling,
+    triplet_text_sampling_with_context_list,
+)
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model_with_BERT import PromptedTransducer
+from optim import Eden, ScaledAdam
+from scaling import Balancer, BiasNorm, Dropout3, ScaleGrad, ScheduledFloat, SwooshR
+from subsampling import Conv2dSubsampling
+from text_normalization import (
+    lower_all_char,
+    lower_only_alpha,
+    train_text_normalization,
+    upper_all_char,
+    upper_only_alpha,
+)
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from zipformer import Zipformer2
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    get_parameter_groups_with_lrs,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+style_transforms = [
+    lambda x: x,  # return it self
+    upper_only_alpha,
+    lower_only_alpha,
+    lower_all_char,
+]
+
+
+def get_first(texts: List[str], pre_texts: List[str]) -> str:
+    out = {
+        "text": texts[0],
+        "pre_text": pre_texts[0],
+        "style_text": "",
+        "transform_ids": 0,
+    }
+    return out
+
+
+def get_upper_only_alpha(texts: List[str], pre_texts: List[str]) -> str:
+    # Always get the first one, which is the gt (mixed-cased trans), but with upper_only_alpha
+    out = {
+        "text": upper_only_alpha(texts[0]),
+        "pre_text": upper_only_alpha(pre_texts[0]),
+        "style_text": "",
+        "transform_ids": 0,
+    }
+    return out
+
+
+def get_adjusted_batch_count(params: AttributeDict) -> float:
+    # returns the number of batches we would have used so far if we had used the reference
+    # duration.  This is for purposes of set_batch_count().
+    return (
+        params.batch_idx_train
+        * (params.max_duration * params.world_size)
+        / params.ref_duration
+    )
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    for name, module in model.named_modules():
+        if hasattr(module, "batch_count"):
+            module.batch_count = batch_count
+        if hasattr(module, "name"):
+            module.name = name
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=str,
+        default="2,2,3,4,3,2",
+        help="Number of zipformer encoder layers per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--downsampling-factor",
+        type=str,
+        default="1,2,4,8,4,2",
+        help="Downsampling factor for each stack of encoder layers.",
+    )
+
+    parser.add_argument(
+        "--feedforward-dim",
+        type=str,
+        default="512,768,1024,1536,1024,768",
+        help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.",
+    )
+
+    parser.add_argument(
+        "--num-heads",
+        type=str,
+        default="4,4,4,8,4,4",
+        help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=str,
+        default="192,256,384,512,384,256",
+        help="Embedding dimension in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--memory-dropout-rate",
+        type=float,
+        default=0.05,
+        help="By which probability, dropout the memory when doing cross-attention.",
+    )
+
+    parser.add_argument(
+        "--memory-layer",
+        type=int,
+        default=0,
+        help="Start doing cross-attention from which layer. Zero-indexed",
+    )
+
+    parser.add_argument(
+        "--query-head-dim",
+        type=str,
+        default="32",
+        help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--value-head-dim",
+        type=str,
+        default="12",
+        help="Value dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-head-dim",
+        type=str,
+        default="4",
+        help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--pos-dim",
+        type=int,
+        default="48",
+        help="Positional-encoding embedding dimension",
+    )
+
+    parser.add_argument(
+        "--encoder-unmasked-dim",
+        type=str,
+        default="192,192,256,256,256,192",
+        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
+        "A single int or comma-separated list.  Must be <= each corresponding encoder_dim.",
+    )
+
+    parser.add_argument(
+        "--cnn-module-kernel",
+        type=str,
+        default="31,31,15,15,15,31",
+        help="Sizes of convolutional kernels in convolution modules in each encoder stack: "
+        "a single int or comma-separated list.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--causal",
+        type=str2bool,
+        default=False,
+        help="If True, use causal version of model.",
+    )
+
+    parser.add_argument(
+        "--chunk-size",
+        type=str,
+        default="16,32,64,-1",
+        help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. "
+        " Must be just -1 if --causal=False",
+    )
+
+    parser.add_argument(
+        "--left-context-frames",
+        type=str,
+        default="64,128,256,-1",
+        help="Maximum left-contexts for causal training, measured in frames which will "
+        "be converted to a number of chunks.  If splitting into chunks, "
+        "chunk left-context frames will be chosen randomly from this list; else not relevant.",
+    )
+
+    parser.add_argument(
+        "--freeze-text-encoder",
+        type=str2bool,
+        default=True,
+    )
+
+    parser.add_argument(
+        "--text-encoder-type",
+        type=str,
+        default="BERT",
+        choices=["BERT", "DistilBERT"],
+        help="Type of the text encoder",
+    )
+
+    parser.add_argument(
+        "--text-encoder-dim",
+        type=int,
+        default=768,
+        help="Dimension of the text encoder",
+    )
+
+    parser.add_argument(
+        "--text-encoder-adapter",
+        type=str2bool,
+        default=False,
+        help="An adapter for pre-trained BERT",
+    )
+
+    parser.add_argument(
+        "--context-injection",
+        type=str2bool,
+        default=False,
+        help="Inject context embedding into the joiner",
+    )
+
+    parser.add_argument(
+        "--context-dropout-rate",
+        type=float,
+        default=0.05,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--base-lr", type=float, default=0.045, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=7500,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=3.5,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--ref-duration",
+        type=float,
+        default=600,
+        help="Reference batch duration for purposes of adjusting batch counts for setting various "
+        "schedules inside the model",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)" "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--use-style-prompt",
+        type=str2bool,
+        default=True,
+        help="Whether to use style prompt.",
+    )
+
+    # arguments for using prompt
+    parser.add_argument(
+        "--pre-text-shuffle-prob",
+        type=float,
+        default=0.05,
+        help="The proportion of pre_text to be shuffled with in a batch",
+    )
+
+    parser.add_argument(
+        "--style-text-shuffle-prob",
+        type=float,
+        default=0.2,
+        help="The proportion of style_text to be shuffled with in a batch",
+    )
+
+    parser.add_argument(
+        "--prompt-mask-prob",
+        type=float,
+        default=0.05,
+        help="The probability of masking prompts",
+    )
+
+    parser.add_argument(
+        "--forced-upper-pre-text",
+        type=str2bool,
+        default=False,
+        help="Forced format of pre-text",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warmup period that dictates the decay of the
+              scale on "simple" (un-pruned) loss.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for zipformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,  # not passed in, this is fixed.
+            "warm_step": 2000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def _to_int_tuple(s: str):
+    return tuple(map(int, s.split(",")))
+
+
+def get_encoder_embed(params: AttributeDict) -> nn.Module:
+    # encoder_embed converts the input of shape (N, T, num_features)
+    # to the shape (N, (T - 7) // 2, encoder_dims).
+    # That is, it does two things simultaneously:
+    #   (1) subsampling: T -> (T - 7) // 2
+    #   (2) embedding: num_features -> encoder_dims
+    # In the normal configuration, we will downsample once more at the end
+    # by a factor of 2, and most of the encoder stacks will run at a lower
+    # sampling rate.
+    encoder_embed = Conv2dSubsampling(
+        in_channels=params.feature_dim,
+        out_channels=_to_int_tuple(params.encoder_dim)[0],
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+    )
+    return encoder_embed
+
+
+class TextEmbedding(nn.Module):
+    def __init__(
+        self,
+        num_embeddings: int = 256,
+        embedding_dim: int = 256,
+        kernel_size: int = 3,
+        layer1_channels: int = 256,
+        layer2_channels: int = 256,
+        bias: bool = True,
+        dropout: float = 0.1,
+    ):
+        super().__init__()
+        self.embed = nn.Embedding(
+            num_embeddings=num_embeddings,  # we encode the text as UTF-8 bytes
+            embedding_dim=embedding_dim,  #
+        )
+
+        assert embedding_dim == layer1_channels  # for depth wise convolution
+        self.conv = nn.Sequential(
+            nn.Conv1d(
+                embedding_dim,
+                layer1_channels,  # depthwise convolution
+                kernel_size=kernel_size,
+                stride=1,
+                padding=(kernel_size - 1) // 2,
+                groups=layer1_channels,
+                bias=True,
+            ),
+            ScaleGrad(0.2),
+            Balancer(layer1_channels, channel_dim=1, min_positive=0.1, max_abs=1.0),
+            nn.ReLU(),
+            nn.Conv1d(
+                layer1_channels,
+                layer2_channels,
+                kernel_size=1,  # pointwise convolution
+                stride=1,
+                padding=0,
+                bias=True,
+            ),
+            Balancer(layer2_channels, channel_dim=1, min_positive=0.1, max_abs=1.0),
+            nn.ReLU(),
+        )
+
+        self.out_norm = BiasNorm(layer2_channels)
+        self.dropout = Dropout3(dropout, shared_dim=1)
+
+    def forward(self, text: torch.Tensor) -> torch.Tensor:
+        """Forward function of the text embedding
+
+        Args:
+            text (torch.Tensor): Text in UTF-8 bytes (T,N)
+        Returns:
+            The embeddings of text (T,N,C)
+        """
+        text = self.embed(text)  # (T,N,C)
+
+        # src = text
+        text = text.permute(1, 2, 0)  # (T,N,C) -> (N,C,T)
+        text = self.conv(text)
+        text = text.permute(2, 0, 1)  # (N,C,T) -> (T,N,C)
+        # src = src + text
+
+        text = self.out_norm(text)
+        text = self.dropout(text)
+
+        return text
+
+
+def get_text_encoder(params: AttributeDict) -> nn.Module:
+    # Return a text encoder
+    if params.text_encoder_type == "BERT":  # This is a BERT-base-cased
+        from transformers import BertModel
+
+        logging.info("Loading pre-trained BERT-base-cased as text encoder")
+        if os.path.exists("data/models/bert-base-cased"):
+            model = BertModel.from_pretrained("data/models/bert-base-cased")
+        else:
+            model = BertModel.from_pretrained("bert-base-cased")
+        assert params.text_encoder_dim == 768
+    elif params.text_encoder_type == "BERT-large":
+        from transformers import BertModel
+
+        logging.info("Loading pre-trained BERT-large-uncased as text encoder")
+        if os.path.exists("data/models/bert-large-uncased"):
+            model = BertModel.from_pretrained("data/models/bert-large-uncased")
+        else:
+            model = BertModel.from_pretrained("bert-large-uncased")
+        assert params.text_encoder_dim == 1024
+    elif params.text_encoder_type == "DistilBERT":
+        from transformers import DistilBertModel  # This is a DistilBERT-base-cased
+
+        logging.info("Loading pre-trained DistilBERT-base-cased as text encoder")
+        model = DistilBertModel.from_pretrained("distilbert-base-cased")
+        assert params.text_encoder_dim == 768
+    else:
+        raise ValueError()
+
+    return model
+
+
+def get_tokenizer(params: AttributeDict):
+
+    if params.text_encoder_type == "BERT":
+        from transformers import BertTokenizer
+
+        # This is a BERT-base-cased
+        if os.path.exists("data/models/bert-base-cased"):
+            tokenizer = BertTokenizer.from_pretrained("data/models/bert-base-cased")
+        else:
+            tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+    elif params.text_encoder_type == "BERT-large":
+        from transformers import BertTokenizer
+
+        # This is a BERT-large-uncased
+        if os.path.exists("data/models/bert-large-uncased"):
+            tokenizer = BertTokenizer.from_pretrained("data/models/bert-large-uncased")
+        else:
+            tokenizer = BertTokenizer.from_pretrained("bert-large-uncased")
+    elif params.text_encoder_type == "DistilBERT":
+        from transformers import DistilBertTokenizer
+
+        tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-cased")
+    else:
+        raise ValueError()
+
+    return tokenizer
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    encoder = Zipformer2(
+        output_downsampling_factor=2,
+        downsampling_factor=_to_int_tuple(params.downsampling_factor),
+        num_encoder_layers=_to_int_tuple(params.num_encoder_layers),
+        encoder_dim=_to_int_tuple(params.encoder_dim),
+        encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim),
+        query_head_dim=_to_int_tuple(params.query_head_dim),
+        pos_head_dim=_to_int_tuple(params.pos_head_dim),
+        value_head_dim=_to_int_tuple(params.value_head_dim),
+        pos_dim=params.pos_dim,
+        num_heads=_to_int_tuple(params.num_heads),
+        feedforward_dim=_to_int_tuple(params.feedforward_dim),
+        cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel),
+        dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)),
+        warmup_batches=4000.0,
+        causal=params.causal,
+        chunk_size=_to_int_tuple(params.chunk_size),
+        left_context_frames=_to_int_tuple(params.left_context_frames),
+        memory_dim=params.text_encoder_dim,  # This is fixed as the BERT base model is 768-D
+        memory_layer=params.memory_layer,
+        memory_dropout_rate=params.memory_dropout_rate,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=max(_to_int_tuple(params.encoder_dim)),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+        context_dim=4 * 768
+        if params.context_injection
+        else -1,  # the output dim of text encoder
+        context_injection=params.context_injection,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    text_encoder = get_text_encoder(params)  # This should be a cased BERT base model
+    num_param = sum([p.numel() for p in text_encoder.parameters()])
+    logging.info(f"Num params in text encoder: {num_param}")
+
+    encoder_embed = get_encoder_embed(params)
+    encoder = get_encoder_model(params)
+
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = PromptedTransducer(
+        encoder_embed=encoder_embed,
+        encoder=encoder,
+        text_encoder=text_encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=int(max(params.encoder_dim.split(","))),
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+        text_encoder_type=params.text_encoder_type,
+        text_encoder_adapter=params.text_encoder_adapter,
+        freeze_text_encoder=params.freeze_text_encoder,
+        context_fuser=None,
+    )
+
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+        if "cur_batch_idx" in saved_params:
+            params["cur_batch_idx"] = saved_params["cur_batch_idx"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def _encode_texts_as_bytes_with_tokenizer(
+    pre_texts: List[str],
+    style_texts: List[str],
+    tokenizer,
+    device: torch.device,
+    max_len: int = 500,
+    no_limit: bool = False,
+) -> Tuple[Dict, Tensor]:
+    """
+    Encode texts as bytes and then integer tensors.
+    Note that the style text will be added to the beginning of texts.
+    """
+    batch_size = len(pre_texts)
+    max_len = min(max_len, 500)
+
+    if no_limit:
+        allowed_lens = [5000 - len(s) for s in style_texts]
+    else:
+        allowed_lens = [1000 - len(s) for s in style_texts]
+    truncated_pre_texts = [pre_texts[i][-allowed_lens[i] :] for i in range(batch_size)]
+    combined_text = [
+        style_texts[i] + " [SEP] " + truncated_pre_texts[i] for i in range(batch_size)
+    ]
+
+    encoded_style_texts = tokenizer(
+        style_texts,
+        return_tensors="pt",
+        padding=True,
+        truncation=True,
+        return_length=True,
+        max_length=max_len,
+    )
+    style_lens = encoded_style_texts["length"].to(device)
+
+    # Use tokenizer to prepare input for text encoder
+    encoded_inputs = tokenizer(
+        combined_text,
+        return_tensors="pt",
+        padding=True,
+        truncation=True,
+        return_length=True,
+        max_length=max_len,
+    ).to(device)
+
+    return encoded_inputs, style_lens
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    tokenizer,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Zipformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    batch_size = feature.size(0)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    warm_step = params.warm_step
+
+    texts = batch["supervisions"]["text"]
+    pre_texts = batch["supervisions"]["pre_text"]
+    style_texts = batch["supervisions"][
+        "style_text"
+    ]  # the style texts are in gt format
+    transform_ids = batch["supervisions"]["transform_ids"]
+
+    # This is to replace full-width symbols with half-width symbols
+    texts = [train_text_normalization(t) for t in texts]
+    pre_texts = [train_text_normalization(t) for t in pre_texts]
+    style_texts = [train_text_normalization(t) for t in style_texts]
+
+    y = sp.encode(
+        texts, out_type=int
+    )  # sp.encode treats consecutive space as a single space
+    y = k2.RaggedTensor(y).to(device)
+
+    if params.forced_upper_pre_text:
+        pre_texts = [upper_only_alpha(p) for p in pre_texts]
+
+    # only shuffle the pre_text and style texts if during training, and use style prompt
+    if is_training:
+        # randomly shuffle&mask the pre_text
+        pre_texts = random_shuffle_subset(
+            pre_texts,
+            p=params.pre_text_shuffle_prob,
+            p_mask=params.prompt_mask_prob,
+        )
+
+        if params.use_style_prompt:
+            if random.random() < 0.5:
+                # randomly shuffle the style_text
+                # now the style_texts are all in gt format
+                style_texts = random_shuffle_subset(
+                    style_texts,
+                    p=params.style_text_shuffle_prob,
+                    p_mask=params.prompt_mask_prob,
+                )
+
+            assert len(transform_ids) == len(style_texts)
+
+            for i in range(len(style_texts)):
+                t = transform_ids[i]  # get the transform id
+                style_texts[i] = style_transforms[t](style_texts[i])
+
+    if not params.use_style_prompt:
+        style_texts = [
+            "" for _ in style_texts
+        ]  # use empty string for style texts if don't use style prompt
+
+    if random.random() < 0.05:
+        logging.info(f"Pre texts: {pre_texts[0]}")
+        logging.info(f"Ref texts: {texts[0]}")
+        logging.info(f"Style texts: {style_texts[0]}")
+
+    encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer(
+        pre_texts=pre_texts,
+        style_texts=style_texts,
+        tokenizer=tokenizer,
+        device=device,
+    )
+
+    if random.random() < 0.02:
+        logging.info(f"Shape of encoded texts: {encoded_inputs['input_ids'].shape} ")
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            encoded_inputs=encoded_inputs,
+            style_lens=style_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+
+        s = params.simple_loss_scale
+        # take down the scale on the simple loss from 1.0 at the start
+        # to params.simple_loss scale by warm_step.
+        simple_loss_scale = (
+            s
+            if batch_idx_train >= warm_step
+            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
+        )
+        pruned_loss_scale = (
+            1.0
+            if batch_idx_train >= warm_step
+            else 0.1 + 0.9 * (batch_idx_train / warm_step)
+        )
+
+        loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    sp: spm.SentencePieceProcessor,
+    tokenizer,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            tokenizer=tokenizer,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    sp: spm.SentencePieceProcessor,
+    tokenizer,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    cur_batch_idx = params.get("cur_batch_idx", 0)
+
+    saved_bad_model = False
+
+    def save_bad_model(suffix: str = ""):
+        save_checkpoint_impl(
+            filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt",
+            model=model,
+            model_avg=model_avg,
+            params=params,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=0,
+        )
+
+    for batch_idx, batch in enumerate(train_dl):
+        if batch_idx % 10 == 0:
+            set_batch_count(model, get_adjusted_batch_count(params))
+        if batch_idx < cur_batch_idx:
+            continue
+        cur_batch_idx = batch_idx
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    tokenizer=tokenizer,
+                    batch=batch,
+                    is_training=True,
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            save_bad_model()
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            params.cur_batch_idx = batch_idx
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            del params.cur_batch_idx
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % 100 == 0 and params.use_fp16:
+            # If the grad scale was less than 1, try increasing it.    The _growth_interval
+            # of the grad scaler is configurable, but we can't configure it to have different
+            # behavior depending on the current grad scale.
+            cur_grad_scale = scaler._scale.item()
+
+            if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0):
+                scaler.update(cur_grad_scale * 2.0)
+            if cur_grad_scale < 0.01:
+                if not saved_bad_model:
+                    save_bad_model(suffix="-first-warning")
+                    saved_bad_model = True
+                logging.warning(f"Grad scale is small: {cur_grad_scale}")
+            if cur_grad_scale < 1.0e-05:
+                save_bad_model()
+                raise RuntimeError(
+                    f"grad_scale is too small, exiting: {cur_grad_scale}"
+                )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = max(scheduler.get_last_lr())
+            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}, "
+                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+                if params.use_fp16:
+                    tb_writer.add_scalar(
+                        "train/grad_scale",
+                        cur_grad_scale,
+                        params.batch_idx_train,
+                    )
+
+        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                tokenizer=tokenizer,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            logging.info(
+                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+            )
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if not params.use_style_prompt:
+        if params.pre_text_shuffle_prob == 0.0:
+            logging.info(
+                f"Pre_text shuffle prob is set to: {params.pre_text_shuffle_prob}"
+            )
+            logging.info(
+                "If style prompt is not used, you should be careful when shuffling the pre_text within the same batch"
+            )
+            logging.info("Hard set this probability to 0.0!")
+            params.pre_text_shuffle_prob = 0.0
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+    tokenizer = get_tokenizer(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model).to(torch.float64)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    if params.freeze_text_encoder:
+        freeze_modules = ["text_encoder"]
+        logging.info(
+            "Freeze the parameters of text encoder and don't include them in the optimizer"
+        )
+    else:
+        freeze_modules = []
+
+    optimizer = ScaledAdam(
+        get_parameter_groups_with_lrs(
+            model, lr=params.base_lr, include_names=True, freeze_modules=freeze_modules
+        ),
+        lr=params.base_lr,  # should have no effect
+        clipping_scale=2.0,
+    )
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        args.max_duration = 100
+        opts = diagnostics.TensorDiagnosticOptions(
+            2**22
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    if params.inf_check:
+        register_inf_check_hooks(model)
+
+    libriheavy = LibriHeavyAsrDataModule(args)
+
+    train_cuts = libriheavy.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        if c.duration < 1.0 or c.duration > 30.0:
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].texts[0], out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].texts[0]}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    if params.use_context_list:
+        text_sampling_func = triplet_text_sampling_with_context_list
+    else:
+        text_sampling_func = triplet_text_sampling
+
+    logging.info(f"Text sampling: {text_sampling_func}")
+
+    train_dl = libriheavy.train_dataloaders(
+        train_cuts,
+        sampler_state_dict=sampler_state_dict,
+        text_sampling_func=text_sampling_func,
+    )
+
+    # For fair comparison, use fixed sampling in valid dataloaders
+    valid_cuts = libriheavy.dev_cuts()
+    valid_dl = libriheavy.valid_dataloaders(
+        valid_cuts, text_sampling_func=naive_triplet_text_sampling
+    )
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            tokenizer=tokenizer,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sp=sp,
+            tokenizer=tokenizer,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    sp: spm.SentencePieceProcessor,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+      sp:
+        The BPE model.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = sp.encode(supervisions["text"], out_type=int)
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    tokenizer: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    sp=sp,
+                    tokenizer=tokenizer,
+                    batch=batch,
+                    is_training=True,
+                )
+            loss.backward()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(batch, params=params, sp=sp)
+            raise
+        logging.info(
+            f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
+        )
+
+
+def main():
+    parser = get_parser()
+    LibriHeavyAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/transcribe_bert.py b/egs/libriheavy/ASR/zipformer_prompt_asr/transcribe_bert.py
new file mode 100644
index 000000000..ef0c48e8a
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/transcribe_bert.py
@@ -0,0 +1,515 @@
+# Copyright    2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+python ./zipformer_prompt_asr/transcribe_bert.py \
+    --epoch 50 \
+    --avg 10 \
+    --exp-dir ./zipformer_prompt_asr/exp \
+    --manifest-dir data/long_audios/long_audio.jsonl.gz \
+    --pre-text-transform mixed-punc \
+    --style-text-transform mixed-punc \
+    --num-history 5 \
+    --use-pre-text True \
+    --use-gt-pre-text False
+
+
+"""
+
+import argparse
+import logging
+import math
+import warnings
+from pathlib import Path
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from decode_bert import _apply_style_transform
+from lhotse import Fbank, load_manifest
+from text_normalization import (
+    lower_all_char,
+    lower_only_alpha,
+    ref_text_normalization,
+    remove_non_alphabetic,
+    train_text_normalization,
+    upper_all_char,
+    upper_only_alpha,
+)
+from tqdm import tqdm
+from train_bert_encoder import (
+    _encode_texts_as_bytes_with_tokenizer,
+    add_model_arguments,
+    get_params,
+    get_tokenizer,
+    get_transducer_model,
+)
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=9,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless7/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+    )
+
+    parser.add_argument(
+        "--manifest-dir",
+        type=str,
+        default="data/long_audios/long_audio.jsonl.gz",
+        help="""This is the manfiest for long audio transcription.
+        The cust are intended to be sorted, i.e first sort by recording ID and
+        then sort by start timestamp""",
+    )
+
+    parser.add_argument(
+        "--use-pre-text",
+        type=str2bool,
+        default=False,
+        help="Whether use pre-text when decoding the current chunk",
+    )
+
+    parser.add_argument(
+        "--use-style-prompt",
+        type=str2bool,
+        default=True,
+        help="Use style prompt when evaluation",
+    )
+
+    parser.add_argument(
+        "--pre-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of content prompt, i.e pre_text",
+    )
+
+    parser.add_argument(
+        "--style-text-transform",
+        type=str,
+        choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"],
+        default="mixed-punc",
+        help="The style of style prompt, i.e style_text",
+    )
+
+    parser.add_argument(
+        "--num-history",
+        type=int,
+        default=2,
+        help="How many previous chunks to look if using pre-text for decoding",
+    )
+
+    parser.add_argument(
+        "--use-gt-pre-text",
+        type=str2bool,
+        default=False,
+        help="Whether use gt pre text when using content prompt",
+    )
+
+    parser.add_argument(
+        "--post-normalization",
+        type=str2bool,
+        default=True,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    params.res_dir = params.exp_dir / "long_audio_transcribe"
+    params.res_dir.mkdir(exist_ok=True)
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "beam_search" in params.method:
+        params.suffix += f"-{params.method}-beam-size-{params.beam_size}"
+
+    if params.use_pre_text:
+        if params.use_gt_pre_text:
+            params.suffix += f"-use-gt-pre-text-{params.pre_text_transform}-history-{params.num_history}"
+        else:
+            params.suffix += (
+                f"-pre-text-{params.pre_text_transform}-history-{params.num_history}"
+            )
+
+    book_name = params.manifest_dir.split("/")[-1].replace(".jsonl.gz", "")
+    setup_logger(
+        f"{params.res_dir}/log-decode-{book_name}-{params.suffix}", log_level="info"
+    )
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+    tokenizer = get_tokenizer(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg + 1
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for" f" --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg + 1:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        filename_start = filenames[-1]
+        filename_end = filenames[0]
+        logging.info(
+            "Calculating the averaged model over iteration checkpoints"
+            f" from {filename_start} (excluded) to {filename_end}"
+        )
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints_with_averaged_model(
+                filename_start=filename_start,
+                filename_end=filename_end,
+                device=device,
+            )
+        )
+    else:
+        assert params.avg > 0, params.avg
+        start = params.epoch - params.avg
+        assert start >= 1, start
+        filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+        filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+        logging.info(
+            f"Calculating the averaged model over epoch range from "
+            f"{start} (excluded) to {params.epoch}"
+        )
+        model.to(device)
+        model.load_state_dict(
+            average_checkpoints_with_averaged_model(
+                filename_start=filename_start,
+                filename_end=filename_end,
+                device=device,
+            )
+        )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    # load manifest
+    manifest = load_manifest(params.manifest_dir)
+
+    results = []
+    count = 0
+
+    last_recording = ""
+    last_end = -1
+    history = []
+    num_pre_texts = []
+
+    for cut in manifest:
+        if cut.has_features:
+            feat = cut.load_features()
+            feat_lens = cut.num_frames
+        else:
+            feat = cut.compute_features(extractor=Fbank())
+            feat_lens = feat.shape[0]
+
+        cur_recording = cut.recording.id
+
+        if cur_recording != last_recording:
+            last_recording = cur_recording
+            history = []  # clean up the history
+            last_end = -1
+            logging.info("Moving on to the next recording")
+        else:
+            if cut.start < last_end - 0.2:  # overlap with the previous cuts
+                logging.warning("An overlap exists between current cut and last cut")
+                logging.warning("Skipping this cut!")
+                continue
+            if cut.start > last_end + 10:
+                logging.warning(
+                    f"Large time gap between the current and previous utterance: {cut.start - last_end}."
+                )
+
+        # prepare input
+        x = torch.tensor(feat, device=device).unsqueeze(0)
+        x_lens = torch.tensor(
+            [
+                feat_lens,
+            ],
+            device=device,
+        )
+
+        if params.use_pre_text:
+            if params.num_history > 0:
+                pre_texts = history[-params.num_history :]
+            else:
+                pre_texts = []
+            num_pre_texts.append(len(pre_texts))
+            pre_texts = [train_text_normalization(" ".join(pre_texts))]
+            fixed_sentence = "Mixed-case English transcription, with punctuation. Actually, it is fully not related."
+            style_texts = [fixed_sentence]
+
+            pre_texts = _apply_style_transform(pre_texts, params.pre_text_transform)
+            if params.use_style_prompt:
+                style_texts = _apply_style_transform(
+                    style_texts, params.style_text_transform
+                )
+
+            # encode prompts
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+
+                encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer(
+                    pre_texts=pre_texts,
+                    style_texts=style_texts,
+                    tokenizer=tokenizer,
+                    device=device,
+                    no_limit=True,
+                )
+                if params.num_history > 5:
+                    logging.info(
+                        f"Shape of encoded texts: {encoded_inputs['input_ids'].shape} "
+                    )
+
+                memory, memory_key_padding_mask = model.encode_text(
+                    encoded_inputs=encoded_inputs,
+                    style_lens=style_lens,
+                )  # (T,B,C)
+        else:
+            memory = None
+            memory_key_padding_mask = None
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            encoder_out, encoder_out_lens = model.encode_audio(
+                feature=x,
+                feature_lens=x_lens,
+                memory=memory,
+                memory_key_padding_mask=memory_key_padding_mask,
+            )
+
+        if params.method == "greedy_search":
+            hyp_tokens = greedy_search_batch(
+                model=model,
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+            )
+        elif params.method == "modified_beam_search":
+            hyp_tokens = modified_beam_search(
+                model=model,
+                encoder_out=encoder_out,
+                encoder_out_lens=encoder_out_lens,
+                beam=params.beam_size,
+            )
+
+        hyp = sp.decode(hyp_tokens)[0]  # in string format
+        ref_text = ref_text_normalization(
+            cut.supervisions[0].texts[0]
+        )  # required to match the training
+
+        # extend the history
+        if params.use_gt_pre_text:
+            history.append(ref_text)
+        else:
+            history.append(hyp)
+        last_end = cut.end  # update the last end timestamp
+
+        # append the current decoding result
+        hyp = hyp.split()
+        ref = ref_text.split()
+        results.append((cut.id, ref, hyp))
+
+        count += 1
+        if count % 100 == 0:
+            logging.info(f"Cuts processed until now: {count}/{len(manifest)}")
+            logging.info(
+                f"Averaged context numbers of last 100 samples is: {sum(num_pre_texts[-100:])/100}"
+            )
+
+    logging.info(f"A total of {count} cuts")
+    logging.info(
+        f"Averaged context numbers of whole set is: {sum(num_pre_texts)/len(num_pre_texts)}"
+    )
+
+    results = sorted(results)
+    recog_path = (
+        params.res_dir / f"recogs-long-audio-{params.method}-{params.suffix}.txt"
+    )
+    store_transcripts(filename=recog_path, texts=results)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    errs_filename = (
+        params.res_dir / f"errs-long-audio-{params.method}-{params.suffix}.txt"
+    )
+    with open(errs_filename, "w") as f:
+        wer = write_error_stats(
+            f,
+            f"long-audio-{params.method}",
+            results,
+            enable_log=True,
+            compute_CER=False,
+        )
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    if params.post_normalization:
+        params.suffix += "-post-normalization"
+
+        new_res = []
+        for item in results:
+            id, ref, hyp = item
+            hyp = upper_only_alpha(" ".join(hyp)).split()
+            ref = upper_only_alpha(" ".join(ref)).split()
+            new_res.append((id, ref, hyp))
+
+    new_res = sorted(new_res)
+    recog_path = (
+        params.res_dir
+        / f"recogs-long-audio-{params.method}-{params.suffix}-post-normalization.txt"
+    )
+    store_transcripts(filename=recog_path, texts=new_res)
+    logging.info(f"The transcripts are stored in {recog_path}")
+
+    errs_filename = (
+        params.res_dir
+        / f"errs-long-audio-{params.method}-{params.suffix}-post-normalization.txt"
+    )
+    with open(errs_filename, "w") as f:
+        wer = write_error_stats(
+            f,
+            f"long-audio-{params.method}",
+            new_res,
+            enable_log=True,
+            compute_CER=False,
+        )
+
+    logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/utils.py b/egs/libriheavy/ASR/zipformer_prompt_asr/utils.py
new file mode 100644
index 000000000..533982519
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/utils.py
@@ -0,0 +1,439 @@
+import argparse
+import ast
+import glob
+import logging
+import os
+from collections import defaultdict
+from typing import Dict, Iterable, List, TextIO, Tuple, Union
+
+import kaldialign
+from lhotse import load_manifest, load_manifest_lazy
+from lhotse.cut import Cut, CutSet
+from text_normalization import remove_non_alphabetic
+from tqdm import tqdm
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--manifest-dir",
+        type=str,
+        default="data/fbank",
+        help="Where are the manifest stored",
+    )
+
+    parser.add_argument(
+        "--subset", type=str, default="medium", help="Which subset to work with"
+    )
+
+    parser.add_argument(
+        "--top-k",
+        type=int,
+        default=10000,
+        help="How many words to keep",
+    )
+
+    return parser
+
+
+def get_facebook_biasing_list(
+    test_set: str,
+    num_distractors: int = 100,
+) -> Dict:
+    # Get the biasing list from the meta paper: https://arxiv.org/pdf/2104.02194.pdf
+    assert num_distractors in (0, 100, 500, 1000, 2000), num_distractors
+    if num_distractors == 0:
+        if test_set == "test-clean":
+            biasing_file = "data/context_biasing/fbai-speech/is21_deep_bias/ref/test-clean.biasing_100.tsv"
+        elif test_set == "test-other":
+            biasing_file = "data/context_biasing/fbai-speech/is21_deep_bias/ref/test-other.biasing_100.tsv"
+        else:
+            raise ValueError(f"Unseen test set {test_set}")
+    else:
+        if test_set == "test-clean":
+            biasing_file = f"data/context_biasing/fbai-speech/is21_deep_bias/ref/test-clean.biasing_{num_distractors}.tsv"
+        elif test_set == "test-other":
+            biasing_file = f"data/context_biasing/fbai-speech/is21_deep_bias/ref/test-other.biasing_{num_distractors}.tsv"
+        else:
+            raise ValueError(f"Unseen test set {test_set}")
+
+    f = open(biasing_file, "r")
+    data = f.readlines()
+    f.close()
+
+    output = dict()
+    for line in data:
+        id, _, l1, l2 = line.split("\t")
+        if num_distractors > 0:  # use distractors
+            biasing_list = ast.literal_eval(l2)
+        else:
+            biasing_list = ast.literal_eval(l1)
+        biasing_list = [w.strip().upper() for w in biasing_list]
+        output[id] = " ".join(biasing_list)
+
+    return output
+
+
+def brian_biasing_list(level: str):
+    # The biasing list from Brian's paper: https://arxiv.org/pdf/2109.00627.pdf
+    root_dir = f"data/context_biasing/LibriSpeechBiasingLists/{level}Level"
+    all_files = glob.glob(root_dir + "/*")
+    biasing_dict = {}
+    for f in all_files:
+        k = f.split("/")[-1]
+        fin = open(f, "r")
+        data = fin.read().strip().split()
+        biasing_dict[k] = " ".join(data)
+        fin.close()
+
+    return biasing_dict
+
+
+def get_rare_words(
+    subset: str = "medium",
+    top_k: int = 10000,
+    # min_count: int = 10000,
+):
+    """Get a list of rare words appearing less than `min_count` times
+
+    Args:
+        subset: The dataset
+        top_k (int): How many frequent words
+    """
+    txt_path = f"data/tmp/transcript_words_{subset}.txt"
+    rare_word_file = f"data/context_biasing/{subset}_rare_words_topk_{top_k}.txt"
+
+    if os.path.exists(rare_word_file):
+        print("File exists, do not proceed!")
+        return
+
+    print("---Identifying rare words in the manifest---")
+    count_file = f"data/tmp/transcript_words_{subset}_count.txt"
+    if not os.path.exists(count_file):
+        with open(txt_path, "r") as file:
+            words = file.read().upper().split()
+            word_count = {}
+            for word in words:
+                word = remove_non_alphabetic(word, strict=False)
+                word = word.split()
+                for w in word:
+                    if w not in word_count:
+                        word_count[w] = 1
+                    else:
+                        word_count[w] += 1
+
+        word_count = list(word_count.items())  # convert to a list of tuple
+        word_count = sorted(word_count, key=lambda w: int(w[1]), reverse=True)
+        with open(count_file, "w") as fout:
+            for w, count in word_count:
+                fout.write(f"{w}\t{count}\n")
+
+    else:
+        word_count = {}
+        with open(count_file, "r") as fin:
+            word_count = fin.read().strip().split("\n")
+            word_count = [pair.split("\t") for pair in word_count]
+            word_count = sorted(word_count, key=lambda w: int(w[1]), reverse=True)
+
+    print(f"A total of {len(word_count)} words appeared!")
+    rare_words = []
+    for word, count in word_count[top_k:]:
+        rare_words.append(word + "\n")
+    print(f"A total of {len(rare_words)} are identified as rare words.")
+
+    with open(rare_word_file, "w") as f:
+        f.writelines(rare_words)
+
+
+def add_context_list_to_manifest(
+    manifest_dir: str,
+    subset: str = "medium",
+    top_k: int = 10000,
+):
+    """Generate a context list of rare words for each utterance in the manifest
+
+    Args:
+        manifest_dir: Where to store the manifest with context list
+        subset (str): Subset
+        top_k (int): How many frequent words
+
+    """
+    orig_manifest_dir = f"{manifest_dir}/libriheavy_cuts_{subset}.jsonl.gz"
+    target_manifest_dir = orig_manifest_dir.replace(
+        ".jsonl.gz", f"_with_context_list_topk_{top_k}.jsonl.gz"
+    )
+    if os.path.exists(target_manifest_dir):
+        print(f"Target file exits at {target_manifest_dir}!")
+        return
+
+    rare_words_file = f"data/context_biasing/{subset}_rare_words_topk_{top_k}.txt"
+    print(f"---Reading rare words from {rare_words_file}---")
+    with open(rare_words_file, "r") as f:
+        rare_words = f.read()
+    rare_words = rare_words.split("\n")
+    rare_words = set(rare_words)
+    print(f"A total of {len(rare_words)} rare words!")
+
+    cuts = load_manifest_lazy(orig_manifest_dir)
+    print(f"Loaded manifest from {orig_manifest_dir}")
+
+    def _add_context(c: Cut):
+        splits = (
+            remove_non_alphabetic(c.supervisions[0].texts[0], strict=False)
+            .upper()
+            .split()
+        )
+        found = []
+        for w in splits:
+            if w in rare_words:
+                found.append(w)
+        c.supervisions[0].context_list = " ".join(found)
+        return c
+
+    cuts = cuts.map(_add_context)
+    print(f"---Saving manifest with context list to {target_manifest_dir}---")
+    cuts.to_file(target_manifest_dir)
+    print("Finished")
+
+
+def check(
+    manifest_dir: str,
+    subset: str = "medium",
+    top_k: int = 10000,
+):
+    # Show how many samples in the training set have a context list
+    # and the average length of context list
+    print("--- Calculating the stats over the manifest ---")
+
+    manifest_dir = f"{manifest_dir}/libriheavy_cuts_{subset}_with_context_list_topk_{top_k}.jsonl.gz"
+    cuts = load_manifest_lazy(manifest_dir)
+    total_cuts = len(cuts)
+    has_context_list = [c.supervisions[0].context_list != "" for c in cuts]
+    context_list_len = [len(c.supervisions[0].context_list.split()) for c in cuts]
+    print(f"{sum(has_context_list)}/{total_cuts} cuts have context list! ")
+    print(
+        f"Average length of non-empty context list is {sum(context_list_len)/sum(has_context_list)}"
+    )
+
+
+def write_error_stats(
+    f: TextIO,
+    test_set_name: str,
+    results: List[Tuple[str, str]],
+    enable_log: bool = True,
+    compute_CER: bool = False,
+    biasing_words: List[str] = None,
+) -> float:
+    """Write statistics based on predicted results and reference transcripts. It also calculates the
+    biasing word error rate as described in https://arxiv.org/pdf/2104.02194.pdf
+
+    It will write the following to the given file:
+
+        - WER
+        - number of insertions, deletions, substitutions, corrects and total
+          reference words. For example::
+
+              Errors: 23 insertions, 57 deletions, 212 substitutions, over 2606
+              reference words (2337 correct)
+
+        - The difference between the reference transcript and predicted result.
+          An instance is given below::
+
+            THE ASSOCIATION OF (EDISON->ADDISON) ILLUMINATING COMPANIES
+
+          The above example shows that the reference word is `EDISON`,
+          but it is predicted to `ADDISON` (a substitution error).
+
+          Another example is::
+
+            FOR THE FIRST DAY (SIR->*) I THINK
+
+          The reference word `SIR` is missing in the predicted
+          results (a deletion error).
+      results:
+        An iterable of tuples. The first element is the cut_id, the second is
+        the reference transcript and the third element is the predicted result.
+      enable_log:
+        If True, also print detailed WER to the console.
+        Otherwise, it is written only to the given file.
+      biasing_words:
+        All the words in the biasing list
+    Returns:
+      Return None.
+    """
+    subs: Dict[Tuple[str, str], int] = defaultdict(int)
+    ins: Dict[str, int] = defaultdict(int)
+    dels: Dict[str, int] = defaultdict(int)
+
+    # `words` stores counts per word, as follows:
+    #   corr, ref_sub, hyp_sub, ins, dels
+    words: Dict[str, List[int]] = defaultdict(lambda: [0, 0, 0, 0, 0])
+    num_corr = 0
+    ERR = "*"
+
+    if compute_CER:
+        for i, res in enumerate(results):
+            cut_id, ref, hyp = res
+            ref = list("".join(ref))
+            hyp = list("".join(hyp))
+            results[i] = (cut_id, ref, hyp)
+
+    for cut_id, ref, hyp in results:
+        ali = kaldialign.align(ref, hyp, ERR)
+        for ref_word, hyp_word in ali:
+            if ref_word == ERR:  # INSERTION
+                ins[hyp_word] += 1
+                words[hyp_word][3] += 1
+            elif hyp_word == ERR:  # DELETION
+                dels[ref_word] += 1
+                words[ref_word][4] += 1
+            elif hyp_word != ref_word:  # SUBSTITUTION
+                subs[(ref_word, hyp_word)] += 1
+                words[ref_word][1] += 1
+                words[hyp_word][2] += 1
+            else:
+                words[ref_word][0] += 1
+                num_corr += 1
+    ref_len = sum([len(r) for _, r, _ in results])
+    sub_errs = sum(subs.values())
+    ins_errs = sum(ins.values())
+    del_errs = sum(dels.values())
+    tot_errs = sub_errs + ins_errs + del_errs
+    tot_err_rate = "%.2f" % (100.0 * tot_errs / ref_len)
+
+    if enable_log:
+        logging.info(
+            f"[{test_set_name}] %WER {tot_errs / ref_len:.2%} "
+            f"[{tot_errs} / {ref_len}, {ins_errs} ins, "
+            f"{del_errs} del, {sub_errs} sub ]"
+        )
+
+    print(f"%WER = {tot_err_rate}", file=f)
+    print(
+        f"Errors: {ins_errs} insertions, {del_errs} deletions, "
+        f"{sub_errs} substitutions, over {ref_len} reference "
+        f"words ({num_corr} correct)",
+        file=f,
+    )
+    print(
+        "Search below for sections starting with PER-UTT DETAILS:, "
+        "SUBSTITUTIONS:, DELETIONS:, INSERTIONS:, PER-WORD STATS:",
+        file=f,
+    )
+
+    print("", file=f)
+    print("PER-UTT DETAILS: corr or (ref->hyp)  ", file=f)
+    for cut_id, ref, hyp in results:
+        ali = kaldialign.align(ref, hyp, ERR)
+        combine_successive_errors = True
+        if combine_successive_errors:
+            ali = [[[x], [y]] for x, y in ali]
+            for i in range(len(ali) - 1):
+                if ali[i][0] != ali[i][1] and ali[i + 1][0] != ali[i + 1][1]:
+                    ali[i + 1][0] = ali[i][0] + ali[i + 1][0]
+                    ali[i + 1][1] = ali[i][1] + ali[i + 1][1]
+                    ali[i] = [[], []]
+            ali = [
+                [
+                    list(filter(lambda a: a != ERR, x)),
+                    list(filter(lambda a: a != ERR, y)),
+                ]
+                for x, y in ali
+            ]
+            ali = list(filter(lambda x: x != [[], []], ali))
+            ali = [
+                [
+                    ERR if x == [] else " ".join(x),
+                    ERR if y == [] else " ".join(y),
+                ]
+                for x, y in ali
+            ]
+
+        print(
+            f"{cut_id}:\t"
+            + " ".join(
+                (
+                    ref_word if ref_word == hyp_word else f"({ref_word}->{hyp_word})"
+                    for ref_word, hyp_word in ali
+                )
+            ),
+            file=f,
+        )
+
+    print("", file=f)
+    print("SUBSTITUTIONS: count ref -> hyp", file=f)
+
+    for count, (ref, hyp) in sorted([(v, k) for k, v in subs.items()], reverse=True):
+        print(f"{count}   {ref} -> {hyp}", file=f)
+
+    print("", file=f)
+    print("DELETIONS: count ref", file=f)
+    for count, ref in sorted([(v, k) for k, v in dels.items()], reverse=True):
+        print(f"{count}   {ref}", file=f)
+
+    print("", file=f)
+    print("INSERTIONS: count hyp", file=f)
+    for count, hyp in sorted([(v, k) for k, v in ins.items()], reverse=True):
+        print(f"{count}   {hyp}", file=f)
+
+    unbiased_word_counts = 0
+    unbiased_word_errs = 0
+    biased_word_counts = 0
+    biased_word_errs = 0
+
+    print("", file=f)
+    print("PER-WORD STATS: word  corr tot_errs count_in_ref count_in_hyp", file=f)
+
+    for _, word, counts in sorted(
+        [(sum(v[1:]), k, v) for k, v in words.items()], reverse=True
+    ):
+        (corr, ref_sub, hyp_sub, ins, dels) = counts
+        tot_errs = ref_sub + hyp_sub + ins + dels
+        # number of appearances of "word" in reference text
+        ref_count = (
+            corr + ref_sub + dels
+        )  # correct + in ref but got substituted + deleted
+        # number of appearances of "word" in hyp text
+        hyp_count = corr + hyp_sub + ins
+
+        if biasing_words is not None:
+            if word in biasing_words:
+                biased_word_counts += ref_count
+                biased_word_errs += ins + dels + ref_sub
+            else:
+                unbiased_word_counts += ref_count
+                unbiased_word_errs += ins + dels + hyp_sub
+
+        print(f"{word}   {corr} {tot_errs} {ref_count} {hyp_count}", file=f)
+
+    if biasing_words is not None:
+        B_WER = "%.2f" % (100 * biased_word_errs / biased_word_counts)
+        U_WER = "%.2f" % (100 * unbiased_word_errs / unbiased_word_counts)
+        logging.info(f"Biased WER: {B_WER} [{biased_word_errs}/{biased_word_counts}] ")
+        logging.info(
+            f"Un-biased WER: {U_WER} [{unbiased_word_errs}/{unbiased_word_counts}]"
+        )
+
+    return float(tot_err_rate)
+
+
+if __name__ == "__main__":
+    parser = get_parser()
+    args = parser.parse_args()
+    manifest_dir = args.manifest_dir
+    subset = args.subset
+    top_k = args.top_k
+    get_rare_words(subset=subset, top_k=top_k)
+    add_context_list_to_manifest(
+        manifest_dir=manifest_dir,
+        subset=subset,
+        top_k=top_k,
+    )
+    check(
+        manifest_dir=manifest_dir,
+        subset=subset,
+        top_k=top_k,
+    )
diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/zipformer.py b/egs/libriheavy/ASR/zipformer_prompt_asr/zipformer.py
new file mode 100644
index 000000000..d1cf90ffb
--- /dev/null
+++ b/egs/libriheavy/ASR/zipformer_prompt_asr/zipformer.py
@@ -0,0 +1,2310 @@
+#!/usr/bin/env python3
+# Copyright    2022-2023  Xiaomi Corp.        (authors: Daniel Povey,
+#                                                       Zengwei Yao,
+#                                                       Xiaoyu Yang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import logging
+import math
+import random
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+from encoder_interface import EncoderInterface
+from scaling import (
+    Identity,  # more friendly to backward hooks than nn.Identity(), for diagnostic reasons.
+)
+from scaling import (
+    ScaledLinear,  # not as in other dirs.. just scales down initial parameter values.
+)
+from scaling import (
+    ActivationDropoutAndLinear,
+    Balancer,
+    BiasNorm,
+    ChunkCausalDepthwiseConv1d,
+    Dropout2,
+    FloatLike,
+    ScheduledFloat,
+    Whiten,
+    convert_num_channels,
+    limit_param_value,
+    penalize_abs_values_gt,
+    softmax,
+)
+from torch import Tensor, nn
+
+
+class Zipformer2(EncoderInterface):
+    """
+    Args:
+
+    Note: all "int or Tuple[int]" arguments below will be treated as lists of the same length
+    as downsampling_factor if they are single ints or one-element tuples.  The length of
+    downsampling_factor defines the number of stacks.
+
+        output_downsampling_factor (int): how much to downsample at the output.  Note:
+            we also downsample by a factor of 2 in the Conv2dSubsampling encoder.
+            You should probably leave this at 2.
+        downsampling_factor (Tuple[int]): downsampling factor for each encoder stack.
+           Note: this is in addition to the downsampling factor of 2 that is applied in
+           the frontend (self.encoder_embed).
+        encoder_dim (Tuple[int]): embedding dimension of each of the encoder stacks, one per
+           encoder stack.
+        num_encoder_layers (int or Tuple[int])): number of encoder layers for each stack
+        encoder_unmasked_dim (int or Tuple[int]): unmasked dimension in each of
+            the encoder stacks for purposes of per-frame dropout (recommend 256 for
+            now).
+        query_head_dim (int or Tuple[int]): dimension of query and key per attention
+           head: per stack, if a tuple..
+        value_head_dim (int or Tuple[int]): dimension of value in each attention head
+        pos_head_dim (int or Tuple[int]): dimension of positional-encoding projection per
+           attention head
+        num_heads: (int or Tuple[int]): number of heads in the self-attention mechanism.
+              Must be at least 4.
+        feedforward_dim (int or Tuple[int]): hidden dimension in feedforward modules
+        cnn_module_kernel (int or Tuple[int])): Kernel size of convolution module
+
+        pos_dim (int): the dimension of each positional-encoding vector prior to projection,
+            e.g. 128.
+
+        dropout (float): dropout rate
+        warmup_batches (float): number of batches to warm up over; this controls
+          dropout of encoder layers.
+        causal (bool): if True, support chunkwise causal convolution.  This should
+          not hurt WER as no modeling power is lost, but the convolution modules will be
+          slightly slower and use more memory.  Enables use of the chunk_size and
+          left_context_chunks options in forward(), which simulates streaming
+          decoding.
+        chunk_size: (list of int): only set this to other than [-1] if causal;
+           the chunk size will be randomly chosen from this list.  -1 means no chunking.
+        left_context_frames: (list of int): determines the number of left-
+           context chunks for causal training; will be rounded to a number of
+           chunks.  Must not be less than cnn_module_kernel (after factoring in
+           rounding and downsampling); an error will be thrown if this is violated.
+        memory_dim: if supplied and >0, will be the dimension of the memory embeddings
+            passed into the zipformer (e.g. this might be the output of another
+            Zipformer used to create embedding vectors.)
+        memory_dropout_rate: By this probability, do not use the provided memory for
+            cross-attention. This should give robustness to the model when evaluated
+            without memory.
+        memory_layer: if supplied and >0, only add cross-attention module starting from
+            the specified layer.
+    """
+
+    def __init__(
+        self,
+        output_downsampling_factor: int = 2,
+        downsampling_factor: Tuple[int] = (2, 4),
+        encoder_dim: Union[int, Tuple[int]] = 384,
+        num_encoder_layers: Union[int, Tuple[int]] = 4,
+        encoder_unmasked_dim: Union[int, Tuple[int]] = 256,
+        query_head_dim: Union[int, Tuple[int]] = 24,
+        pos_head_dim: Union[int, Tuple[int]] = 4,
+        value_head_dim: Union[int, Tuple[int]] = 12,
+        num_heads: Union[int, Tuple[int]] = 8,
+        feedforward_dim: Union[int, Tuple[int]] = 1536,
+        cnn_module_kernel: Union[int, Tuple[int]] = 31,
+        pos_dim: int = 192,
+        dropout: FloatLike = None,  # see code below for default
+        warmup_batches: float = 4000.0,
+        causal: bool = False,
+        chunk_size: Tuple[int] = [-1],
+        left_context_frames: Tuple[int] = [-1],
+        memory_dim: int = -1,
+        memory_dropout_rate: float = 0.05,
+        memory_layer: int = 0,
+    ) -> None:
+        super(Zipformer2, self).__init__()
+
+        if dropout is None:
+            dropout = ScheduledFloat((0.0, 0.3), (20000.0, 0.1))
+
+        def _to_tuple(x):
+            """Converts a single int or a 1-tuple of an int to a tuple with the same length
+            as downsampling_factor"""
+            if isinstance(x, int):
+                x = (x,)
+            if len(x) == 1:
+                x = x * len(downsampling_factor)
+            else:
+                assert len(x) == len(downsampling_factor) and isinstance(x[0], int)
+            return x
+
+        self.output_downsampling_factor = output_downsampling_factor  # int
+        self.downsampling_factor = downsampling_factor  # tuple
+        self.encoder_dim = encoder_dim = _to_tuple(encoder_dim)  # tuple
+        self.encoder_unmasked_dim = encoder_unmasked_dim = _to_tuple(
+            encoder_unmasked_dim
+        )  # tuple
+        num_encoder_layers = _to_tuple(num_encoder_layers)
+        self.num_encoder_layers = num_encoder_layers
+        self.query_head_dim = query_head_dim = _to_tuple(query_head_dim)
+        self.value_head_dim = value_head_dim = _to_tuple(value_head_dim)
+        pos_head_dim = _to_tuple(pos_head_dim)
+        self.num_heads = num_heads = _to_tuple(num_heads)
+        feedforward_dim = _to_tuple(feedforward_dim)
+        self.cnn_module_kernel = cnn_module_kernel = _to_tuple(cnn_module_kernel)
+
+        self.causal = causal
+        self.chunk_size = chunk_size
+        self.left_context_frames = left_context_frames
+        self.memory_dropout_rate = memory_dropout_rate
+        self.memory_layer = memory_layer
+
+        for u, d in zip(encoder_unmasked_dim, encoder_dim):
+            assert u <= d
+
+        # each one will be Zipformer2Encoder or DownsampledZipformer2Encoder
+        encoders = []
+
+        num_encoders = len(downsampling_factor)
+        for i in range(num_encoders):
+            encoder_layer = Zipformer2EncoderLayer(
+                embed_dim=encoder_dim[i],
+                pos_dim=pos_dim,
+                num_heads=num_heads[i],
+                query_head_dim=query_head_dim[i],
+                pos_head_dim=pos_head_dim[i],
+                value_head_dim=value_head_dim[i],
+                feedforward_dim=feedforward_dim[i],
+                memory_dim=memory_dim if i >= self.memory_layer else -1,
+                dropout=dropout,
+                cnn_module_kernel=cnn_module_kernel[i],
+                causal=causal,
+            )
+
+            # For the segment of the warmup period, we let the Conv2dSubsampling
+            # layer learn something.  Then we start to warm up the other encoders.
+            encoder = Zipformer2Encoder(
+                encoder_layer,
+                num_encoder_layers[i],
+                pos_dim=pos_dim,
+                dropout=dropout,
+                warmup_begin=warmup_batches * (i + 1) / (num_encoders + 1),
+                warmup_end=warmup_batches * (i + 2) / (num_encoders + 1),
+                final_layerdrop_rate=0.035 * (downsampling_factor[i] ** 0.5),
+            )
+
+            if downsampling_factor[i] != 1:
+                encoder = DownsampledZipformer2Encoder(
+                    encoder,
+                    dim=encoder_dim[i],
+                    downsample=downsampling_factor[i],
+                    dropout=dropout,
+                )
+
+            encoders.append(encoder)
+
+        self.encoders = nn.ModuleList(encoders)
+
+        self.downsample_output = SimpleDownsample(
+            max(encoder_dim), downsample=output_downsampling_factor, dropout=dropout
+        )
+
+    def get_feature_masks(self, x: Tensor) -> Union[List[float], List[Tensor]]:
+        """
+        In eval mode, returns [1.0] * num_encoders; in training mode, returns a number of
+        randomized feature masks, one per encoder.
+        On e.g. 15% of frames, these masks will zero out all enocder dims larger than
+        some supplied number, e.g. >256, so in effect on those frames we are using
+        a smaller encoer dim.
+
+        We generate the random masks at this level because we want the 2 masks to 'agree'
+        all the way up the encoder stack. This will mean that the 1st mask will have
+        mask values repeated self.zipformer_subsampling_factor times.
+
+        Args:
+           x: the embeddings (needed for the shape and dtype and device), of shape
+             (1, batch_size, encoder_dims0)
+        """
+        num_encoders = len(self.encoder_dim)
+        if not self.training:
+            return [1.0] * num_encoders
+
+        (num_frames0, batch_size, _encoder_dims0) = x.shape
+
+        assert self.encoder_dim[0] == _encoder_dims0, (
+            self.encoder_dim[0],
+            _encoder_dims0,
+        )
+
+        feature_mask_dropout_prob = 0.125
+
+        # mask1 shape: (1, batch_size, 1)
+        mask1 = (
+            torch.rand(1, batch_size, 1, device=x.device) > feature_mask_dropout_prob
+        ).to(x.dtype)
+
+        # mask2 has additional sequences masked, about twice the number.
+        mask2 = torch.logical_and(
+            mask1,
+            (
+                torch.rand(1, batch_size, 1, device=x.device)
+                > feature_mask_dropout_prob
+            ).to(x.dtype),
+        )
+
+        # dim: (1, batch_size, 2)
+        mask = torch.cat((mask1, mask2), dim=-1)
+
+        feature_masks = []
+        for i in range(num_encoders):
+            channels = self.encoder_dim[i]
+            feature_mask = torch.ones(
+                1, batch_size, channels, dtype=x.dtype, device=x.device
+            )
+            u1 = self.encoder_unmasked_dim[i]
+            u2 = u1 + (channels - u1) // 2
+
+            feature_mask[:, :, u1:u2] *= mask[..., 0:1]
+            feature_mask[:, :, u2:] *= mask[..., 1:2]
+
+            feature_masks.append(feature_mask)
+
+        return feature_masks
+
+    def get_chunk_info(self) -> Tuple[int, int]:
+        """
+        Returns chunk_size and left_context_chunks.
+        """
+        if not self.causal:
+            return -1, -1
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            assert len(self.chunk_size) == 1, self.chunk_size
+            chunk_size = self.chunk_size[0]
+        else:
+            chunk_size = random.choice(self.chunk_size)
+        if chunk_size == -1:
+            left_context_chunks = -1
+        else:
+            if torch.jit.is_scripting() or torch.jit.is_tracing():
+                assert len(self.left_context_frames) == 1, self.left_context_frames
+                left_context_frames = self.left_context_frames[0]
+            else:
+                left_context_frames = random.choice(self.left_context_frames)
+            # Note: in Python, -1 // n == -1 for n > 0
+            left_context_chunks = left_context_frames // chunk_size
+            if left_context_chunks == 0:
+                left_context_chunks = 1
+
+        return chunk_size, left_context_chunks
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        memory: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          src_key_padding_mask:
+            The mask for padding, of shape (batch_size, seq_len); True means
+             masked position. May be None.
+          memory:  optionally, the memory embeddings of shape (memory_len, batch_size, memory_dim)
+          memory_key_padding_mask: optionally the mask for padding of memory input (for source-
+             attention), of shape  (batch_size, memory_len); True means
+              masked position.  May be None.
+
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (batch_size, output_seq_len, max(encoder_dim))
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+        """
+        outputs = []
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            feature_masks = [1.0] * len(self.encoder_dim)
+        else:
+            feature_masks = self.get_feature_masks(x)
+
+        chunk_size, left_context_chunks = self.get_chunk_info()
+
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            # Not support exporting a model for simulating streaming decoding
+            attn_mask = None
+        else:
+            attn_mask = self._get_attn_mask(x, chunk_size, left_context_chunks)
+
+        if self.training and memory is not None:
+            batch_size = x.shape[1]
+            # setting memory to zero should be equivalent to not using the
+            # memory input at all, since the Attention module has no biases.
+            memory = memory * (
+                torch.rand(batch_size, 1, device=memory.device)
+                > self.memory_dropout_rate
+            )
+
+        for i, module in enumerate(self.encoders):
+            ds = self.downsampling_factor[i]
+            x = convert_num_channels(x, self.encoder_dim[i])
+
+            x = module(
+                x,
+                chunk_size=chunk_size,
+                feature_mask=feature_masks[i],
+                src_key_padding_mask=(
+                    None
+                    if src_key_padding_mask is None
+                    else src_key_padding_mask[..., ::ds]
+                ),
+                attn_mask=attn_mask,
+                memory=memory if i >= self.memory_layer else None,
+                memory_key_padding_mask=memory_key_padding_mask
+                if i >= self.memory_layer
+                else None,
+            )
+            outputs.append(x)
+
+        # if the last output has the largest dimension, x will be unchanged,
+        # it will be the same as outputs[-1].  Otherwise it will be concatenated
+        # from different pieces of 'outputs', taking each dimension from the
+        # most recent output that has it present.
+        x = self._get_full_dim_output(outputs)
+        x = self.downsample_output(x)
+        # class Downsample has this rounding behavior..
+        assert self.output_downsampling_factor == 2, self.output_downsampling_factor
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            lengths = (x_lens + 1) // 2
+        else:
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                lengths = (x_lens + 1) // 2
+
+        return x, lengths
+
+    def _get_attn_mask(
+        self, x: Tensor, chunk_size: int, left_context_chunks: int
+    ) -> Optional[Tensor]:
+        """
+        Return None if chunk_size == -1, else return attention mask of shape
+          (seq_len, seq_len), interpreted as (tgt_seq_len, src_seq_len).  True
+           means a masked position.
+        Args:
+           x: embeddings after self.encoder_embed(), of shape (seq_len, batch_size, embed_dim).
+          chunk_size: chunk size, must divide
+        """
+        if chunk_size <= 0:
+            return None
+        assert all(chunk_size % d == 0 for d in self.downsampling_factor)
+        if left_context_chunks >= 0:
+            num_encoders = len(self.encoder_dim)
+            assert all(
+                chunk_size * left_context_chunks
+                >= (self.cnn_module_kernel[i] // 2) * self.downsampling_factor[i]
+                for i in range(num_encoders)
+            )
+        else:
+            left_context_chunks = 1000000
+
+        seq_len = x.shape[0]
+
+        # t is frame index, shape (seq_len,)
+        t = torch.arange(seq_len, dtype=torch.int32, device=x.device)
+        # c is chunk index for each frame, shape (seq_len,)
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            c = t // chunk_size
+        else:
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                c = t // chunk_size
+        src_c = c
+        tgt_c = c.unsqueeze(-1)
+
+        attn_mask = torch.logical_or(src_c > tgt_c, src_c < tgt_c - left_context_chunks)
+        if __name__ == "__main__":
+            logging.info(f"attn_mask = {attn_mask}")
+        return attn_mask
+
+    def _get_full_dim_output(self, outputs: List[Tensor]):
+        num_encoders = len(self.encoder_dim)
+        assert len(outputs) == num_encoders
+        output_dim = max(self.encoder_dim)
+        output_pieces = [outputs[-1]]
+        cur_dim = self.encoder_dim[-1]
+        for i in range(num_encoders - 2, -1, -1):
+            d = self.encoder_dim[i]
+            if d > cur_dim:
+                this_output = outputs[i]
+                output_pieces.append(this_output[..., cur_dim:d])
+                cur_dim = d
+        assert cur_dim == output_dim
+        return torch.cat(output_pieces, dim=-1)
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        x_lens: Tensor,
+        states: List[Tensor],
+        src_key_padding_mask: Tensor,
+    ) -> Tuple[Tensor, Tensor, List[Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (seq_len, batch_size, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+          states: list of cached tensors of all encoder layers. For layer-i,
+            states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2,
+            cached_conv1, cached_conv2).
+          src_key_padding_mask:
+            The mask for padding, of shape (batch_size, seq_len); True means
+            masked position. May be None.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - embeddings: its shape is (output_seq_len, batch_size, max(encoder_dim))
+            - lengths, a tensor of shape (batch_size,) containing the number
+              of frames in `embeddings` before padding.
+            - updated states
+        """
+        outputs = []
+        new_states = []
+        layer_offset = 0
+
+        for i, module in enumerate(self.encoders):
+            num_layers = module.num_layers
+            ds = self.downsampling_factor[i]
+            x = convert_num_channels(x, self.encoder_dim[i])
+
+            x, new_layer_states = module.streaming_forward(
+                x,
+                states=states[layer_offset * 6 : (layer_offset + num_layers) * 6],
+                left_context_len=self.left_context_frames[0] // ds,
+                src_key_padding_mask=src_key_padding_mask[..., ::ds],
+            )
+            layer_offset += num_layers
+            outputs.append(x)
+            new_states += new_layer_states
+
+        # if the last output has the largest dimension, x will be unchanged,
+        # it will be the same as outputs[-1].  Otherwise it will be concatenated
+        # from different pieces of 'outputs', taking each dimension from the
+        # most recent output that has it present.
+        x = self._get_full_dim_output(outputs)
+        x = self.downsample_output(x)
+        # class Downsample has this rounding behavior..
+        assert self.output_downsampling_factor == 2
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            lengths = (x_lens + 1) // 2
+        else:
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                lengths = (x_lens + 1) // 2
+
+        return x, lengths, new_states
+
+    @torch.jit.export
+    def get_init_states(
+        self,
+        batch_size: int = 1,
+        device: torch.device = torch.device("cpu"),
+    ) -> List[Tensor]:
+        """Get initial states.
+
+        A list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6]
+        is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2).
+        """
+        states = []
+        for i, module in enumerate(self.encoders):
+            num_layers = module.num_layers
+            embed_dim = self.encoder_dim[i]
+            ds = self.downsampling_factor[i]
+            num_heads = self.num_heads[i]
+            key_dim = self.query_head_dim[i] * num_heads
+            value_dim = self.value_head_dim[i] * num_heads
+            downsample_left = self.left_context_frames[0] // ds
+            nonlin_attn_head_dim = 3 * embed_dim // 4
+            conv_left_pad = self.cnn_module_kernel[i] // 2
+            for layer in range(num_layers):
+                cached_key = torch.zeros(downsample_left, batch_size, key_dim).to(
+                    device
+                )
+                cached_nonlin_attn = torch.zeros(
+                    1, batch_size, downsample_left, nonlin_attn_head_dim
+                ).to(device)
+                cached_val1 = torch.zeros(downsample_left, batch_size, value_dim).to(
+                    device
+                )
+                cached_val2 = torch.zeros(downsample_left, batch_size, value_dim).to(
+                    device
+                )
+                cached_conv1 = torch.zeros(batch_size, embed_dim, conv_left_pad).to(
+                    device
+                )
+                cached_conv2 = torch.zeros(batch_size, embed_dim, conv_left_pad).to(
+                    device
+                )
+                states += [
+                    cached_key,
+                    cached_nonlin_attn,
+                    cached_val1,
+                    cached_val2,
+                    cached_conv1,
+                    cached_conv2,
+                ]
+
+        return states
+
+
+def _whitening_schedule(x: float, ratio: float = 2.0) -> ScheduledFloat:
+    return ScheduledFloat((0.0, x), (20000.0, ratio * x), default=x)
+
+
+def _balancer_schedule(min_prob: float):
+    return ScheduledFloat((0.0, 0.4), (8000.0, min_prob))
+
+
+class Zipformer2EncoderLayer(nn.Module):
+    """
+    Args:
+        embed_dim: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        feedforward_dim: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+
+    Examples::
+        >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        pos_dim: int,
+        num_heads: int,
+        query_head_dim: int,
+        pos_head_dim: int,
+        value_head_dim: int,
+        feedforward_dim: int,
+        dropout: FloatLike = 0.1,
+        cnn_module_kernel: int = 31,
+        causal: bool = False,
+        memory_dim: int = -1,
+        attention_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0
+        ),
+        conv_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0
+        ),
+        const_attention_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.25), (4000.0, 0.025), default=0
+        ),
+        ff2_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0)
+        ),
+        ff3_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0)
+        ),
+        bypass_skip_rate: FloatLike = ScheduledFloat(
+            (0.0, 0.5), (4000.0, 0.02), default=0
+        ),
+    ) -> None:
+        super(Zipformer2EncoderLayer, self).__init__()
+        self.embed_dim = embed_dim
+
+        # self.bypass implements layer skipping as well as bypass; see its default values.
+        self.bypass = BypassModule(
+            embed_dim, skip_rate=bypass_skip_rate, straight_through_rate=0
+        )
+        # bypass_mid is bypass used in the middle of the layer.
+        self.bypass_mid = BypassModule(embed_dim, straight_through_rate=0)
+
+        # skip probability for dynamic modules (meaning: anything but feedforward).
+        self.attention_skip_rate = copy.deepcopy(attention_skip_rate)
+        # an additional skip probability that applies to ConvModule to stop it from
+        # contributing too much early on.
+        self.conv_skip_rate = copy.deepcopy(conv_skip_rate)
+
+        # ff2_skip_rate is to prevent the ff2 module from having output that's too big
+        # compared to its residual.
+        self.ff2_skip_rate = copy.deepcopy(ff2_skip_rate)
+        self.ff3_skip_rate = copy.deepcopy(ff3_skip_rate)
+
+        self.const_attention_rate = copy.deepcopy(const_attention_rate)
+
+        self.self_attn_weights = RelPositionMultiheadAttentionWeights(
+            embed_dim,
+            pos_dim=pos_dim,
+            num_heads=num_heads,
+            query_head_dim=query_head_dim,
+            pos_head_dim=pos_head_dim,
+            dropout=0.0,
+        )
+
+        self.self_attn1 = Attention(embed_dim, embed_dim, num_heads, value_head_dim)
+
+        self.self_attn2 = Attention(embed_dim, embed_dim, num_heads, value_head_dim)
+
+        if memory_dim > 0:
+            self.attn_weights = MultiheadAttentionWeights(
+                memory_dim,
+                embed_dim,
+                num_heads=num_heads,
+                head_dim=query_head_dim,
+                dropout=0.0,
+            )
+            self.src_attn1 = Attention(memory_dim, embed_dim, num_heads, value_head_dim)
+            self.src_attn2 = Attention(memory_dim, embed_dim, num_heads, value_head_dim)
+            self.memory_balancer = Balancer(
+                embed_dim,
+                channel_dim=-1,
+                min_abs=0.015,
+            )
+
+        self.feed_forward1 = FeedforwardModule(
+            embed_dim, (feedforward_dim * 3) // 4, dropout
+        )
+
+        self.feed_forward2 = FeedforwardModule(embed_dim, feedforward_dim, dropout)
+
+        self.feed_forward3 = FeedforwardModule(
+            embed_dim, (feedforward_dim * 5) // 4, dropout
+        )
+
+        self.nonlin_attention = NonlinAttention(
+            embed_dim, hidden_channels=3 * embed_dim // 4
+        )
+
+        self.conv_module1 = ConvolutionModule(
+            embed_dim, cnn_module_kernel, causal=causal
+        )
+
+        self.conv_module2 = ConvolutionModule(
+            embed_dim, cnn_module_kernel, causal=causal
+        )
+
+        # self.attention_squeeze = AttentionSqueeze(embed_dim, embed_dim // 2)
+
+        self.norm = BiasNorm(embed_dim)
+
+        self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5))
+
+        self.balancer1 = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.45,
+            max_positive=0.55,
+            min_abs=0.2,
+            max_abs=4.0,
+        )
+
+        # balancer for output of NonlinAttentionModule
+        self.balancer_na = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.3,
+            max_positive=0.7,
+            min_abs=ScheduledFloat((0.0, 0.004), (4000.0, 0.02)),
+            prob=0.05,  # out of concern for memory usage
+        )
+
+        # balancer for output of feedforward2, prevent it from staying too
+        # small.  give this a very small probability, even at the start of
+        # training, it's to fix a rare problem and it's OK to fix it slowly.
+        self.balancer_ff2 = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.3,
+            max_positive=0.7,
+            min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.1), default=0.0),
+            max_abs=2.0,
+            prob=0.05,
+        )
+
+        self.balancer_ff3 = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.3,
+            max_positive=0.7,
+            min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.2), default=0.0),
+            max_abs=4.0,
+            prob=0.05,
+        )
+
+        self.whiten = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(4.0, ratio=3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+        self.balancer2 = Balancer(
+            embed_dim,
+            channel_dim=-1,
+            min_positive=0.45,
+            max_positive=0.55,
+            min_abs=0.1,
+            max_abs=4.0,
+        )
+
+    def get_bypass_scale(self, batch_size: int):
+        # returns bypass-scale of shape (num_channels,),
+        # or (batch_size, num_channels,).  This is actually the
+        # scale on the non-residual term, so 0 correponds to bypassing
+        # this module.
+        if torch.jit.is_scripting() or not self.training:
+            return self.bypass_scale
+        else:
+            ans = limit_param_value(
+                self.bypass_scale,
+                min=float(self.bypass_min),
+                max=float(self.bypass_max),
+            )
+            layer_skip_rate = float(self.layer_skip_rate)
+            if layer_skip_rate != 0.0:
+                mask = torch.rand((batch_size, 1), device=ans.device) > layer_skip_rate
+                ans = ans * mask
+                # now ans is of shape (batch_size, num_channels), and is zero for sequences
+                # on which we have randomly chosen to do layer-skipping.
+            return ans
+
+    def get_sequence_dropout_mask(
+        self, x: Tensor, dropout_rate: float
+    ) -> Optional[Tensor]:
+        if dropout_rate == 0.0 or not self.training or torch.jit.is_scripting():
+            return None
+        batch_size = x.shape[1]
+        mask = (torch.rand(batch_size, 1, device=x.device) > dropout_rate).to(x.dtype)
+        return mask
+
+    def sequence_dropout(self, x: Tensor, dropout_rate: float) -> Tensor:
+        """
+        Apply sequence-level dropout to x.
+        x shape: (seq_len, batch_size, embed_dim)
+        """
+        dropout_mask = self.get_sequence_dropout_mask(x, dropout_rate)
+        if dropout_mask is None:
+            return x
+        else:
+            return x * dropout_mask
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        chunk_size: int = -1,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        memory: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """
+            Pass the input through the encoder layer.
+            Args:
+                src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim).
+             pos_emb: (1, 2*seq_len-1, pos_emb_dim) or (batch_size, 2*seq_len-1, pos_emb_dim)
+             chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking.
+           feature_mask: something that broadcasts with src, that we'll multiply `src`
+                  by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim)
+             attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len),
+                    interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len).
+                   True means masked position. May be None.
+        src_key_padding_mask:  the mask for padding, of shape (batch_size, seq_len); True means
+                 masked position.  May be None.
+
+            Returns:
+               A tensor which has the same shape as src
+        """
+        src_orig = src
+
+        # dropout rate for non-feedforward submodules
+        attention_skip_rate = float(self.attention_skip_rate) if self.training else 0.0
+
+        # attn_weights: (num_heads, batch_size, seq_len, seq_len)
+        attn_weights = self.self_attn_weights(
+            src,
+            pos_emb=pos_emb,
+            attn_mask=attn_mask,
+            key_padding_mask=src_key_padding_mask,
+        )
+
+        if memory is not None and hasattr(self, "attn_weights"):
+            src_attn_weights = self.attn_weights(memory, src, memory_key_padding_mask)
+
+        src = src + self.feed_forward1(src)
+
+        attn_dropout_mask = self.get_sequence_dropout_mask(src, attention_skip_rate)
+
+        if True:
+            selected_attn_weights = attn_weights[0:2]
+            if random.random() < float(self.const_attention_rate):
+                # Make attention weights constant.  The intention is to
+                # encourage these modules to do something similar to an
+                # averaging-over-time operation.
+                # only need the mask, can just use the 1st one and expand later
+                selected_attn_weights = selected_attn_weights[0:1]
+                selected_attn_weights = (selected_attn_weights > 0.0).to(
+                    selected_attn_weights.dtype
+                )
+                selected_attn_weights = selected_attn_weights * (
+                    1.0 / selected_attn_weights.sum(dim=-1, keepdim=True)
+                )
+                selected_attn_weights = selected_attn_weights.expand(2, -1, -1, -1)
+
+        na = self.balancer_na(self.nonlin_attention(src, selected_attn_weights[0:1]))
+
+        src = src + (na if attn_dropout_mask is None else na * attn_dropout_mask)
+
+        self_attn = self.self_attn1(src, attn_weights)
+
+        src = src + (
+            self_attn if attn_dropout_mask is None else self_attn * attn_dropout_mask
+        )
+
+        if memory is not None and hasattr(self, "attn_weights"):
+            src = src + self.sequence_dropout(
+                self.memory_balancer(self.src_attn1(memory, src_attn_weights)),
+                attention_skip_rate,
+            )
+
+        src = src + self.sequence_dropout(
+            self.conv_module1(
+                src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask
+            ),
+            float(self.conv_skip_rate),
+        )
+
+        src = src + self.sequence_dropout(
+            self.balancer_ff2(self.feed_forward2(src)), float(self.ff2_skip_rate)
+        )
+
+        # bypass in the middle of the layer.
+        src = self.bypass_mid(src_orig, src)
+
+        self_attn = self.self_attn2(src, attn_weights)
+
+        src = src + (
+            self_attn if attn_dropout_mask is None else self_attn * attn_dropout_mask
+        )
+
+        if memory is not None and hasattr(self, "attn_weights"):
+            src = src + self.sequence_dropout(
+                self.memory_balancer(self.src_attn2(memory, src_attn_weights)),
+                attention_skip_rate,
+            )
+
+        src = src + self.sequence_dropout(
+            self.conv_module2(
+                src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask
+            ),
+            float(self.conv_skip_rate),
+        )
+
+        src = src + self.sequence_dropout(
+            self.balancer_ff3(self.feed_forward3(src)), float(self.ff3_skip_rate)
+        )
+
+        src = self.balancer1(src)
+        src = self.norm(src)
+
+        src = self.bypass(src_orig, src)
+
+        src = self.balancer2(src)
+        src = self.whiten(src)
+
+        return src
+
+
+class Zipformer2Encoder(nn.Module):
+    r"""Zipformer2Encoder is a stack of N encoder layers
+
+    Args:
+     encoder_layer: an instance of the Zipformer2EncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+           pos_dim: the dimension for the relative positional encoding
+
+    Examples::
+        >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8)
+        >>> zipformer_encoder = Zipformer2Encoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = zipformer_encoder(src)
+    """
+
+    def __init__(
+        self,
+        encoder_layer: nn.Module,
+        num_layers: int,
+        pos_dim: int,
+        dropout: float,
+        warmup_begin: float,
+        warmup_end: float,
+        initial_layerdrop_rate: float = 0.5,
+        final_layerdrop_rate: float = 0.05,
+    ) -> None:
+        super().__init__()
+        self.encoder_pos = CompactRelPositionalEncoding(
+            pos_dim, dropout_rate=0.15, length_factor=1.0
+        )
+
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+        assert 0 <= warmup_begin <= warmup_end
+
+        delta = (1.0 / num_layers) * (warmup_end - warmup_begin)
+        cur_begin = warmup_begin  # interpreted as a training batch index
+        for i in range(num_layers):
+            cur_end = cur_begin + delta
+            self.layers[i].bypass.skip_rate = ScheduledFloat(
+                (cur_begin, initial_layerdrop_rate),
+                (cur_end, final_layerdrop_rate),
+                default=0.0,
+            )
+            cur_begin = cur_end
+
+    def forward(
+        self,
+        src: Tensor,
+        chunk_size: int = -1,
+        feature_mask: Union[Tensor, float] = 1.0,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        memory: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim).
+            chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking.
+            feature_mask: something that broadcasts with src, that we'll multiply `src`
+               by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim)
+            attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len),
+                 interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len).
+                 True means masked position. May be None.
+            src_key_padding_mask:  the mask for padding, of shape (batch_size, seq_len); True means
+                 masked position.  May be None.
+            memory:  optionally, the memory embeddings of shape (memory_len, batch_size, memory_dim)
+            memory_key_padding_mask: optionally the mask for padding of memory input (for source-
+                attention), of shape  (batch_size, memory_len); True means
+                 masked position.  May be None.
+
+        Returns: a Tensor with the same shape as src.
+        """
+        pos_emb = self.encoder_pos(src)
+        output = src
+
+        output = output * feature_mask
+
+        for i, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                chunk_size=chunk_size,
+                attn_mask=attn_mask,
+                src_key_padding_mask=src_key_padding_mask,
+                memory=memory,
+                memory_key_padding_mask=memory_key_padding_mask,
+            )
+
+            output = output * feature_mask
+
+        return output
+
+
+class BypassModule(nn.Module):
+    """
+    An nn.Module that implements a learnable bypass scale, and also randomized per-sequence
+    layer-skipping.  The bypass is limited during early stages of training to be close to
+    "straight-through", i.e. to not do the bypass operation much initially, in order to
+    force all the modules to learn something.
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        skip_rate: FloatLike = 0.0,
+        straight_through_rate: FloatLike = 0.0,
+        scale_min: FloatLike = ScheduledFloat((0.0, 0.9), (20000.0, 0.2), default=0),
+        scale_max: FloatLike = 1.0,
+    ):
+        super().__init__()
+        self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5))
+        self.skip_rate = copy.deepcopy(skip_rate)
+        self.straight_through_rate = copy.deepcopy(straight_through_rate)
+        self.scale_min = copy.deepcopy(scale_min)
+        self.scale_max = copy.deepcopy(scale_max)
+
+    def _get_bypass_scale(self, batch_size: int):
+        # returns bypass-scale of shape (num_channels,),
+        # or (batch_size, num_channels,).  This is actually the
+        # scale on the non-residual term, so 0 correponds to bypassing
+        # this module.
+        if torch.jit.is_scripting() or not self.training:
+            return self.bypass_scale
+        else:
+            ans = limit_param_value(
+                self.bypass_scale, min=float(self.scale_min), max=float(self.scale_max)
+            )
+            skip_rate = float(self.skip_rate)
+            if skip_rate != 0.0:
+                mask = torch.rand((batch_size, 1), device=ans.device) > skip_rate
+                ans = ans * mask
+                # now ans is of shape (batch_size, num_channels), and is zero for sequences
+                # on which we have randomly chosen to do layer-skipping.
+            straight_through_rate = float(self.straight_through_rate)
+            if straight_through_rate != 0.0:
+                mask = (
+                    torch.rand((batch_size, 1), device=ans.device)
+                    < straight_through_rate
+                )
+                ans = torch.maximum(ans, mask.to(ans.dtype))
+
+            return ans
+
+    def forward(self, src_orig: Tensor, src: Tensor):
+        """
+        Args: src_orig and src are both of shape (seq_len, batch_size, num_channels)
+        Returns: something with the same shape as src and src_orig
+        """
+        bypass_scale = self._get_bypass_scale(src.shape[1])
+        return src_orig + (src - src_orig) * bypass_scale
+
+
+class DownsampledZipformer2Encoder(nn.Module):
+    r"""
+    DownsampledZipformer2Encoder is a zipformer encoder evaluated at a reduced frame rate,
+    after convolutional downsampling, and then upsampled again at the output, and combined
+    with the origin input, so that the output has the same shape as the input.
+    """
+
+    def __init__(
+        self, encoder: nn.Module, dim: int, downsample: int, dropout: FloatLike
+    ):
+        super(DownsampledZipformer2Encoder, self).__init__()
+        self.downsample_factor = downsample
+        self.downsample = SimpleDownsample(dim, downsample, dropout)
+        self.encoder = encoder
+        self.upsample = SimpleUpsample(dim, downsample)
+        self.out_combiner = BypassModule(dim, straight_through_rate=0.025)
+
+    def forward(
+        self,
+        src: Tensor,
+        chunk_size: int = -1,
+        feature_mask: Union[Tensor, float] = 1.0,
+        attn_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        memory: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Tensor]:
+        r"""Downsample, go through encoder, upsample.
+
+        Args:
+            src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim).
+            feature_mask: something that broadcasts with src, that we'll multiply `src`
+               by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim)
+            attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len),
+                 interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len).
+                 True means masked position. May be None.
+            src_key_padding_mask:  the mask for padding, of shape (batch_size, seq_len); True means
+                 masked position.  May be None.
+            memory:  optionally, the memory embeddings of shape (memory_len, batch_size, memory_dim)
+            memory_key_padding_mask: optionally the mask for padding of memory input (for source-
+                attention), of shape  (batch_size, memory_len); True means
+                 masked position.  May be None.
+
+        Returns: a Tensor with the same shape as src.
+        """
+        src_orig = src
+        src = self.downsample(src)
+        ds = self.downsample_factor
+        if attn_mask is not None:
+            attn_mask = attn_mask[::ds, ::ds]
+
+        src = self.encoder(
+            src,
+            chunk_size=chunk_size // ds,
+            feature_mask=feature_mask,
+            attn_mask=attn_mask,
+            src_key_padding_mask=src_key_padding_mask,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )
+        src = self.upsample(src)
+        # remove any extra frames that are not a multiple of downsample_factor
+        src = src[: src_orig.shape[0]]
+
+        return self.out_combiner(src_orig, src)
+
+
+class SimpleDownsample(torch.nn.Module):
+    """
+    Does downsampling with attention, by weighted sum, and a projection..
+    """
+
+    def __init__(self, channels: int, downsample: int, dropout: FloatLike):
+        super(SimpleDownsample, self).__init__()
+
+        self.bias = nn.Parameter(torch.zeros(downsample))
+
+        self.name = None  # will be set from training code
+        self.dropout = copy.deepcopy(dropout)
+
+        self.downsample = downsample
+
+    def forward(self, src: Tensor) -> Tensor:
+        """
+        x: (seq_len, batch_size, in_channels)
+        Returns a tensor of shape
+           ( (seq_len+downsample-1)//downsample, batch_size, channels)
+        """
+        (seq_len, batch_size, in_channels) = src.shape
+        ds = self.downsample
+        d_seq_len = (seq_len + ds - 1) // ds
+
+        # Pad to an exact multiple of self.downsample
+        if seq_len != d_seq_len * ds:
+            # right-pad src, repeating the last element.
+            pad = d_seq_len * ds - seq_len
+            src_extra = src[src.shape[0] - 1 :].expand(pad, src.shape[1], src.shape[2])
+            src = torch.cat((src, src_extra), dim=0)
+            assert src.shape[0] == d_seq_len * ds
+
+        src = src.reshape(d_seq_len, ds, batch_size, in_channels)
+
+        weights = self.bias.softmax(dim=0)
+        # weights: (downsample, 1, 1)
+        weights = weights.unsqueeze(-1).unsqueeze(-1)
+
+        # ans1 is the first `in_channels` channels of the output
+        ans = (src * weights).sum(dim=1)
+
+        return ans
+
+
+class SimpleUpsample(torch.nn.Module):
+    """
+    A very simple form of upsampling that mostly just repeats the input, but
+    also adds a position-specific bias.
+    """
+
+    def __init__(self, num_channels: int, upsample: int):
+        super(SimpleUpsample, self).__init__()
+        self.upsample = upsample
+
+    def forward(self, src: Tensor) -> Tensor:
+        """
+        x: (seq_len, batch_size, num_channels)
+        Returns a tensor of shape
+           ( (seq_len*upsample), batch_size, num_channels)
+        """
+        upsample = self.upsample
+        (seq_len, batch_size, num_channels) = src.shape
+        src = src.unsqueeze(1).expand(seq_len, upsample, batch_size, num_channels)
+        src = src.reshape(seq_len * upsample, batch_size, num_channels)
+        return src
+
+
+class CompactRelPositionalEncoding(torch.nn.Module):
+    """
+    Relative positional encoding module.  This version is "compact" meaning it is able to encode
+    the important information about the relative position in a relatively small number of dimensions.
+    The goal is to make it so that small differences between large relative offsets (e.g. 1000 vs. 1001)
+    make very little difference to the embedding.   Such differences were potentially important
+    when encoding absolute position, but not important when encoding relative position because there
+    is now no need to compare two large offsets with each other.
+
+    Our embedding works done by projecting the interval [-infinity,infinity] to a finite interval
+    using the atan() function, before doing the fourier transform of that fixed interval.  The
+    atan() function would compress the "long tails" too small,
+    making it hard to distinguish between different magnitudes of large offsets, so we use a logarithmic
+    function to compress large offsets to a smaller range before applying atan().
+    Scalings are chosen in such a way that the embedding can clearly distinguish invidual offsets as long
+    as they are quite close to the origin, e.g. abs(offset) <= about sqrt(embedding_dim)
+
+
+    Args:
+        embed_dim: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length: just a heuristic for initialization.
+        length_factor: a heuristic scale (should be >= 1.0) which, if larger, gives
+           less weight to small differences of offset near the origin.
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        dropout_rate: FloatLike,
+        max_len: int = 1000,
+        length_factor: float = 1.0,
+    ) -> None:
+        """Construct a CompactRelPositionalEncoding object."""
+        super(CompactRelPositionalEncoding, self).__init__()
+        self.embed_dim = embed_dim
+        assert embed_dim % 2 == 0
+        self.dropout = Dropout2(dropout_rate)
+        self.pe = None
+        assert length_factor >= 1.0
+        self.length_factor = length_factor
+        self.extend_pe(torch.tensor(0.0).expand(max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(0) >= x.size(0) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+
+        T = x.size(0)
+        # if T == 4, x would contain [ -3, -2, 1, 0, 1, 2, 3 ]
+        x = torch.arange(-(T - 1), T, device=x.device).to(torch.float32).unsqueeze(1)
+
+        freqs = 1 + torch.arange(self.embed_dim // 2, device=x.device)
+
+        # `compression_length` this is arbitrary/heuristic, if it is larger we have more resolution
+        # for small time offsets but less resolution for large time offsets.
+        compression_length = self.embed_dim**0.5
+        # x_compressed, like X, goes from -infinity to infinity as T goes from -infinity to infinity;
+        # but it does so more slowly than T for large absolute values of T.
+        # The formula is chosen so that d(x_compressed )/dx is 1 around x == 0, which
+        # is important.
+        x_compressed = (
+            compression_length
+            * x.sign()
+            * ((x.abs() + compression_length).log() - math.log(compression_length))
+        )
+
+        # if self.length_factor == 1.0, then length_scale is chosen so that the
+        # FFT can exactly separate points close to the origin (T == 0).  So this
+        # part of the formulation is not really heuristic.
+        # But empirically, for ASR at least, length_factor > 1.0 seems to work better.
+        length_scale = self.length_factor * self.embed_dim / (2.0 * math.pi)
+
+        # note for machine implementations: if atan is not available, we can use:
+        #   x.sign() * ((1 / (x.abs() + 1)) - 1)  * (-math.pi/2)
+        #  check on wolframalpha.com: plot(sign(x) *  (1 / ( abs(x) + 1) - 1 ) * -pi/2 , atan(x))
+        x_atan = (x_compressed / length_scale).atan()  # results between -pi and pi
+
+        cosines = (x_atan * freqs).cos()
+        sines = (x_atan * freqs).sin()
+
+        pe = torch.zeros(x.shape[0], self.embed_dim, device=x.device)
+        pe[:, 0::2] = cosines
+        pe[:, 1::2] = sines
+        pe[:, -1] = 1.0  # for bias.
+
+        self.pe = pe.to(dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tensor:
+        """Create positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (time, batch, `*`).
+
+        Returns:
+            positional embedding, of shape (1, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        pos_emb = self.pe[
+            self.pe.size(0) // 2
+            - x.size(0)
+            + 1 : self.pe.size(0) // 2  # noqa E203
+            + x.size(0),
+            :,
+        ]
+        pos_emb = pos_emb.unsqueeze(0)
+        return self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttentionWeights(nn.Module):
+    r"""Module that computes multi-head attention weights with relative position encoding.
+    Various other modules consume the resulting attention weights: see, for example, the
+    SimpleAttention module which allows you to compute conventional attention.
+
+    This is a quite heavily modified from: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context",
+    we have to write up the differences.
+
+
+    Args:
+           embed_dim: number of channels at the input to this module, e.g. 256
+             pos_dim: dimension of the positional encoding vectors, e.g. 128.
+           num_heads:  number of heads to compute weights for, e.g. 8
+     query_head_dim: dimension of the query (and key), per head.  e.g. 24.
+       pos_head_dim: dimension of the projected positional encoding per head, e.g. 4.
+            dropout: dropout probability for attn_output_weights. Default: 0.0.
+       pos_emb_skip_rate: probability for skipping the pos_emb part of the scores on
+                     any given call to forward(), in training time.
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        pos_dim: int,
+        num_heads: int,
+        query_head_dim: int,
+        pos_head_dim: int,
+        dropout: float = 0.0,
+        pos_emb_skip_rate: FloatLike = ScheduledFloat((0.0, 0.5), (4000.0, 0.0)),
+    ) -> None:
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.query_head_dim = query_head_dim
+        self.pos_head_dim = pos_head_dim
+        self.dropout = dropout
+        self.pos_emb_skip_rate = copy.deepcopy(pos_emb_skip_rate)
+        self.name = None  # will be overwritten in training code; for diagnostics.
+
+        key_head_dim = query_head_dim
+        in_proj_dim = (query_head_dim + key_head_dim + pos_head_dim) * num_heads
+
+        # the initial_scale is supposed to take over the "scaling" factor of
+        # head_dim ** -0.5 that has been used in previous forms of attention,
+        # dividing it between the query and key.   Note: this module is intended
+        # to be used with the ScaledAdam optimizer; with most other optimizers,
+        # it would be necessary to apply the scaling factor in the forward function.
+        self.in_proj = ScaledLinear(
+            embed_dim, in_proj_dim, bias=True, initial_scale=query_head_dim**-0.25
+        )
+
+        self.whiten_keys = Whiten(
+            num_groups=num_heads,
+            whitening_limit=_whitening_schedule(3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+
+        # add a balancer for the keys that runs with very small probability, and
+        # tries to enforce that all dimensions have mean around zero.  The
+        # weights produced by this module are invariant to adding a constant to
+        # the keys, so the derivative of the bias is mathematically zero; but
+        # due to how Adam/ScaledAdam work, it can learn a fairly large nonzero
+        # bias because the small numerical roundoff tends to have a non-random
+        # sign.  This module is intended to prevent that.  Use a very small
+        # probability; that should be suffixient to fix the problem.
+        self.balance_keys = Balancer(
+            key_head_dim * num_heads,
+            channel_dim=-1,
+            min_positive=0.4,
+            max_positive=0.6,
+            min_abs=0.0,
+            max_abs=100.0,
+            prob=0.025,
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(
+            pos_dim, num_heads * pos_head_dim, bias=False, initial_scale=0.05
+        )
+
+        # the following are for diagnosics only, see --print-diagnostics option
+        self.copy_pos_query = Identity()
+        self.copy_query = Identity()
+
+    def forward(
+        self,
+        x: Tensor,
+        pos_emb: Tensor,
+        chunk_size: int = -1,
+        key_padding_mask: Optional[Tensor] = None,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""
+        Args:
+            x: input of shape (seq_len, batch_size, embed_dim)
+            pos_emb: Positional embedding tensor, of shape (1, 2*seq_len - 2, pos_dim)
+           chunk_size
+            key_padding_mask: a bool tensor of shape (batch_size, seq_len).  Positions that
+               are True in this mask will be ignored as sources in the attention weighting.
+            attn_mask: mask of shape (seq_len, seq_len) or (batch_size, seq_len, seq_len),
+               interpreted as ([batch_size,] tgt_seq_len, src_seq_len)
+               saying which positions are allowed to attend to which other positions.
+        Returns:
+           a tensor of attention weights, of shape (hum_heads, batch_size, seq_len, seq_len)
+           interpreted as (hum_heads, batch_size, tgt_seq_len, src_seq_len).
+        """
+        x = self.in_proj(x)
+        query_head_dim = self.query_head_dim
+        pos_head_dim = self.pos_head_dim
+        num_heads = self.num_heads
+
+        seq_len, batch_size, _ = x.shape
+
+        query_dim = query_head_dim * num_heads
+
+        q = x[..., 0:query_dim]
+        k = x[..., query_dim : 2 * query_dim]
+        # p is the position-encoding query
+        p = x[..., 2 * query_dim :]
+        assert p.shape[-1] == num_heads * pos_head_dim
+
+        q = self.copy_query(q)  # for diagnostics only, does nothing.
+        k = self.whiten_keys(self.balance_keys(k))  # does nothing in the forward pass.
+        p = self.copy_pos_query(p)  # for diagnostics only, does nothing.
+
+        q = q.reshape(seq_len, batch_size, num_heads, query_head_dim)
+        p = p.reshape(seq_len, batch_size, num_heads, pos_head_dim)
+        k = k.reshape(seq_len, batch_size, num_heads, query_head_dim)
+
+        # time1 refers to target, time2 refers to source.
+        q = q.permute(2, 1, 0, 3)  # (head, batch, time1, query_head_dim)
+        p = p.permute(2, 1, 0, 3)  # (head, batch, time1, pos_head_dim)
+        k = k.permute(2, 1, 3, 0)  # (head, batch, d_k, time2)
+
+        attn_scores = torch.matmul(q, k)
+
+        if not self.training or random.random() >= float(self.pos_emb_skip_rate):
+            pos_emb = self.linear_pos(pos_emb)
+            seq_len2 = 2 * seq_len - 1
+            pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute(
+                2, 0, 3, 1
+            )
+            # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2)
+
+            # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2)
+            #  [where seq_len2 represents relative position.]
+            pos_scores = torch.matmul(p, pos_emb)
+            # the following .as_strided() expression converts the last axis of pos_scores from relative
+            # to absolute position.  I don't know whether I might have got the time-offsets backwards or
+            # not, but let this code define which way round it is supposed to be.
+            pos_scores = pos_scores.as_strided(
+                (num_heads, batch_size, seq_len, seq_len),
+                (
+                    pos_scores.stride(0),
+                    pos_scores.stride(1),
+                    pos_scores.stride(2) - pos_scores.stride(3),
+                    pos_scores.stride(3),
+                ),
+                storage_offset=pos_scores.stride(3) * (seq_len - 1),
+            )
+
+            attn_scores = attn_scores + pos_scores
+
+        if self.training and random.random() < 0.1:
+            # This is away of limiting the attention scores to not be
+            # too large.  It incurs a penalty if any of them has an absolute
+            # value greater than 25.0.  this should be outside the normal range
+            # of the attention scores.  We use this mechanism instead of, say,
+            # something added to the loss function involving the entropy,
+            # because once the entropy gets very small gradients through the
+            # softmax can become very small, and we'd get zero derivatives.  The
+            # choices of 1.0e-04 as the scale on the penalty makes this
+            # mechanism vulnerable to the absolute scale of the loss function,
+            # but we view this as a failsafe to avoid "implausible" parameter
+            # values rather than a regularization method that should be active
+            # under normal circumstances.
+            attn_scores = penalize_abs_values_gt(
+                attn_scores, limit=25.0, penalty=1.0e-04, name=self.name
+            )
+
+        assert attn_scores.shape == (num_heads, batch_size, seq_len, seq_len)
+
+        if attn_mask is not None:
+            assert attn_mask.dtype == torch.bool
+            # use -1000 to avoid nan's where attn_mask and key_padding_mask make
+            # all scores zero.  It's important that this be large enough that exp(-1000)
+            # is exactly zero, for reasons related to const_attention_rate, it
+            # compares the final weights with zero.
+            attn_scores = attn_scores.masked_fill(attn_mask, -1000)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.shape == (
+                batch_size,
+                seq_len,
+            ), key_padding_mask.shape
+            attn_scores = attn_scores.masked_fill(
+                key_padding_mask.unsqueeze(1),
+                -1000,
+            )
+
+        # We use our own version of softmax, defined in scaling.py, which should
+        # save a little of the memory used in backprop by, if we are in
+        # automatic mixed precision mode (amp / autocast), by only storing the
+        # half-precision output for backprop purposes.
+        attn_weights = softmax(attn_scores, dim=-1)
+
+        if random.random() < 0.001:
+            self._print_attn_entropy(attn_weights)
+
+        attn_weights = nn.functional.dropout(
+            attn_weights, p=self.dropout, training=self.training
+        )
+
+        return attn_weights
+
+    def _print_attn_entropy(self, attn_weights: Tensor):
+        # attn_weights: (num_heads, batch_size, seq_len, seq_len)
+        (num_heads, batch_size, seq_len, seq_len) = attn_weights.shape
+
+        with torch.no_grad():
+            with torch.cuda.amp.autocast(enabled=False):
+                attn_weights = attn_weights.to(torch.float32)
+                attn_weights_entropy = (
+                    -((attn_weights + 1.0e-20).log() * attn_weights)
+                    .sum(dim=-1)
+                    .mean(dim=(1, 2))
+                )
+                logging.info(
+                    f"name={self.name}, attn_weights_entropy = {attn_weights_entropy}"
+                )
+
+
+class Attention(nn.Module):
+    """
+    The simplest possible attention module.  This one works with already-computed attention
+    weights, e.g. as computed by RelPositionMultiheadAttentionWeights.
+
+    Args:
+          embed_dim_in: the input embedding dimension
+          embed_dim_out: the output embedding dimension (normally the same as input)
+          num_heads: the number of attention heads
+          value_head_dim: the value dimension per head
+    """
+
+    def __init__(
+        self,
+        embed_dim_in: int,
+        embed_dim_out: int,
+        num_heads: int,
+        value_head_dim: int,
+    ) -> None:
+        super().__init__()
+        self.in_proj = nn.Linear(embed_dim_in, num_heads * value_head_dim, bias=False)
+
+        # Note we set bias to False so that input of 0 will have no effect
+        self.out_proj = ScaledLinear(
+            num_heads * value_head_dim, embed_dim_out, bias=False, initial_scale=0.05
+        )
+
+        self.whiten = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(7.5, ratio=3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+    ) -> Tensor:
+        """
+        Args:
+          x: input tensor, of shape (seq_len, batch_size, embed_dim)
+         attn_weights: a tensor of shape (num_heads, batch_size, query_len, key_len),
+          Expect attn_weights.sum(dim=-1) == 1. The input here is the value in the
+          original attention mechanism.
+        Returns:
+           a tensor with the same shape as x.
+        """
+        (num_heads, batch_size, query_len, key_len) = attn_weights.shape
+
+        x = self.in_proj(x)  # (key_len, batch_size, num_heads * value_head_dim)
+        x = x.reshape(key_len, batch_size, num_heads, -1).permute(2, 1, 0, 3)
+        # now x: (num_heads, batch_size, key_len, value_head_dim)
+        value_head_dim = x.shape[-1]
+
+        # todo: see whether there is benefit in overriding matmul
+        x = torch.matmul(attn_weights, x)
+        # v: (num_heads, batch_size, query_len, value_head_dim)
+
+        x = (
+            x.permute(2, 1, 0, 3)
+            .contiguous()
+            .view(query_len, batch_size, num_heads * value_head_dim)
+        )
+
+        # returned value is of shape (query_len, batch_size, embed_dim), like the input.
+        x = self.out_proj(x)
+        x = self.whiten(x)
+
+        return x
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+        cached_val: Tensor,
+        left_context_len: int,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+            x: input tensor, of shape (seq_len, batch_size, embed_dim)
+            attn_weights: a tensor of shape (num_heads, batch_size, seq_len, seq_len),
+              with seq_len being interpreted as (tgt_seq_len, src_seq_len).  Expect
+              attn_weights.sum(dim=-1) == 1.
+            cached_val: cached attention value tensor of left context,
+              of shape (left_context_len, batch_size, value_dim)
+            left_context_len: number of left context frames.
+
+        Returns:
+           - attention weighted output, a tensor with the same shape as x.
+           - updated cached attention value tensor of left context.
+        """
+        (seq_len, batch_size, embed_dim) = x.shape
+        num_heads = attn_weights.shape[0]
+        seq_len2 = seq_len + left_context_len
+        assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len2)
+
+        x = self.in_proj(x)  # (seq_len, batch_size, num_heads * value_head_dim)
+
+        # Pad cached left contexts
+        assert cached_val.shape[0] == left_context_len, (
+            cached_val.shape[0],
+            left_context_len,
+        )
+        x = torch.cat([cached_val, x], dim=0)
+        # Update cached left contexts
+        cached_val = x[-left_context_len:, ...]
+
+        x = x.reshape(seq_len2, batch_size, num_heads, -1).permute(2, 1, 0, 3)
+        # now x: (num_heads, batch_size, seq_len, value_head_dim)
+        value_head_dim = x.shape[-1]
+
+        # todo: see whether there is benefit in overriding matmul
+        x = torch.matmul(attn_weights, x)
+        # v: (num_heads, batch_size, seq_len, value_head_dim)
+
+        x = (
+            x.permute(2, 1, 0, 3)
+            .contiguous()
+            .view(seq_len, batch_size, num_heads * value_head_dim)
+        )
+
+        # returned value is of shape (seq_len, batch_size, embed_dim), like the input.
+        x = self.out_proj(x)
+
+        return x, cached_val
+
+
+class MultiheadAttentionWeights(nn.Module):
+    r"""Module that computes multi-head cross-attention weights.  Allows src and target
+    to have different dims.
+
+    Args:
+          key_embed_dim: number of channels of the thing that we'll project to
+              make the query (corresponds to source).  e.g. 256
+          query_embed_dim: number of channels of the thing that we'll project to
+              make the query (corresponds to target).  e.g. 256
+          num_heads:  number of heads to compute weights for, e.g. 8
+           head_dim: dimension of the query and key, per head.  e.g. 24.
+             dropout: dropout probability for attn_output_weights. Default: 0.0.
+    """
+
+    def __init__(
+        self,
+        key_embed_dim: int,
+        query_embed_dim: int,
+        num_heads: int,
+        head_dim: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super().__init__()
+        self.key_embed_dim = key_embed_dim
+        self.query_embed_dim = query_embed_dim
+        self.num_heads = num_heads
+        self.head_dim = head_dim
+        self.dropout = dropout
+        self.name = None  # will be overwritten in training code; for diagnostics.
+
+        # the initial_scale is supposed to take over the "scaling" factor of
+        # head_dim ** -0.5 that has been used in previous forms of attention,
+        # dividing it between the query and key.   Note: this module is intended
+        # to be used with the ScaledAdam optimizer; with most other optimizers,
+        # it would be necessary to apply the scaling factor in the forward function.
+        self.query_in_proj = ScaledLinear(
+            query_embed_dim,
+            head_dim * num_heads,
+            bias=True,
+            initial_scale=head_dim**-0.25,
+        )
+
+        # weights produced by this module are invariant to adding a constant to
+        # the keys, so we don't need a bias for the keys.
+        self.key_in_proj = ScaledLinear(
+            key_embed_dim,
+            head_dim * num_heads,
+            bias=False,
+            initial_scale=head_dim**-0.25,
+        )
+
+        self.whiten_keys = Whiten(
+            num_groups=num_heads,
+            whitening_limit=_whitening_schedule(3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.025,
+        )
+
+    def forward(
+        self,
+        key: Tensor,
+        query: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""
+        Args:
+              key: input of shape (key_len, batch_size, key_embed_dim)
+            query: input of shape (query_len, batch_size, query_embed_dim)
+          key_padding_mask: an optional bool tensor of shape (batch_size, key_len).  Positions that
+               are True in this mask will be ignored as sources in the attention weighting.
+        Returns:
+           a tensor of attention weights, of shape (hum_heads, batch_size, query_len, key_len)
+        """
+        q = self.query_in_proj(query)
+        k = self.key_in_proj(key)
+
+        head_dim = self.head_dim
+        num_heads = self.num_heads
+
+        query_len, batch_size, _ = q.shape
+        key_len, _batch_size, _ = k.shape
+        assert _batch_size == batch_size
+
+        k = self.whiten_keys(k)  # does nothing in the forward pass.
+
+        q = q.reshape(query_len, batch_size, num_heads, head_dim)
+        k = k.reshape(key_len, batch_size, num_heads, head_dim)
+
+        # time1 refers to target, time2 refers to source.
+        q = q.permute(2, 1, 0, 3)  # (head, batch, time1, query_head_dim)
+        k = k.permute(2, 1, 3, 0)  # (head, batch, d_k, time2)
+
+        attn_scores = torch.matmul(q, k)
+
+        if self.training and random.random() < 0.1:
+            # This is a way of limiting the attention scores to not be
+            # too large.  It incurs a penalty if any of them has an absolute
+            # value greater than 25.0.  this should be outside the normal range
+            # of the attention scores.  We use this mechanism instead of, say,
+            # something added to the loss function involving the entropy,
+            # because once the entropy gets very small gradients through the
+            # softmax can become very small, and we'd get zero derivatives.  The
+            # choices of 1.0e-04 as the scale on the penalty makes this
+            # mechanism vulnerable to the absolute scale of the loss function,
+            # but we view this as a failsafe to avoid "implausible" parameter
+            # values rather than a regularization method that should be active
+            # under normal circumstances.
+            attn_scores = penalize_abs_values_gt(
+                attn_scores, limit=25.0, penalty=1.0e-04, name=self.name
+            )
+
+        assert attn_scores.shape == (num_heads, batch_size, query_len, key_len)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.shape == (
+                batch_size,
+                key_len,
+            ), key_padding_mask.shape
+            attn_scores = attn_scores.masked_fill(
+                key_padding_mask.unsqueeze(1),
+                -1000,
+            )
+
+        # We use our own version of softmax, defined in scaling.py, which should
+        # save a little of the memory used in backprop by, if we are in
+        # automatic mixed precision mode (amp / autocast), by only storing the
+        # half-precision output for backprop purposes.
+        attn_weights = softmax(attn_scores, dim=-1)
+
+        if random.random() < 0.001:
+            self._print_attn_entropy(attn_weights)
+
+        attn_weights = nn.functional.dropout(
+            attn_weights, p=self.dropout, training=self.training
+        )
+
+        return attn_weights
+
+    def _print_attn_entropy(self, attn_weights: Tensor):
+        # attn_weights: (num_heads, batch_size, seq_len, seq_len)
+        (num_heads, batch_size, seq_len, seq_len) = attn_weights.shape
+
+        with torch.no_grad():
+            with torch.cuda.amp.autocast(enabled=False):
+                attn_weights = attn_weights.to(torch.float32)
+                attn_weights_entropy = (
+                    -((attn_weights + 1.0e-20).log() * attn_weights)
+                    .sum(dim=-1)
+                    .mean(dim=(1, 2))
+                )
+                logging.info(
+                    f"name={self.name}, attn_weights_entropy = {attn_weights_entropy}"
+                )
+
+
+class FeedforwardModule(nn.Module):
+    """Feedforward module in Zipformer2 model."""
+
+    def __init__(self, embed_dim: int, feedforward_dim: int, dropout: FloatLike):
+        super(FeedforwardModule, self).__init__()
+        self.in_proj = nn.Linear(embed_dim, feedforward_dim)
+
+        self.hidden_balancer = Balancer(
+            feedforward_dim,
+            channel_dim=-1,
+            min_positive=0.3,
+            max_positive=1.0,
+            min_abs=0.75,
+            max_abs=5.0,
+        )
+
+        # shared_dim=0 means we share the dropout mask along the time axis
+        self.out_proj = ActivationDropoutAndLinear(
+            feedforward_dim,
+            embed_dim,
+            activation="SwooshL",
+            dropout_p=dropout,
+            dropout_shared_dim=0,
+            bias=True,
+            initial_scale=0.1,
+        )
+
+        self.out_whiten = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(7.5),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+    def forward(self, x: Tensor):
+        x = self.in_proj(x)
+        x = self.hidden_balancer(x)
+        # out_proj contains SwooshL activation, then dropout, then linear.
+        x = self.out_proj(x)
+        x = self.out_whiten(x)
+        return x
+
+
+class NonlinAttention(nn.Module):
+    """This is like the ConvolutionModule, but refactored so that we use multiplication by attention weights (borrowed
+       from the attention module) in place of actual convolution.  We also took out the second nonlinearity, the
+       one after the attention mechanism.
+
+    Args:
+        channels (int): The number of channels of conv layers.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        hidden_channels: int,
+    ) -> None:
+        super().__init__()
+
+        self.hidden_channels = hidden_channels
+
+        self.in_proj = nn.Linear(channels, hidden_channels * 3, bias=True)
+
+        # balancer that goes before the sigmoid.  Have quite a large min_abs value, at 2.0,
+        # because we noticed that well-trained instances of this module have abs-value before the sigmoid
+        # starting from about 3, and poorly-trained instances of the module have smaller abs values
+        # before the sigmoid.
+        self.balancer = Balancer(
+            hidden_channels,
+            channel_dim=-1,
+            min_positive=ScheduledFloat((0.0, 0.25), (20000.0, 0.05)),
+            max_positive=ScheduledFloat((0.0, 0.75), (20000.0, 0.95)),
+            min_abs=0.5,
+            max_abs=5.0,
+        )
+        self.tanh = nn.Tanh()
+
+        self.identity1 = Identity()  # for diagnostics.
+        self.identity2 = Identity()  # for diagnostics.
+        self.identity3 = Identity()  # for diagnostics.
+
+        self.out_proj = ScaledLinear(
+            hidden_channels, channels, bias=True, initial_scale=0.05
+        )
+
+        self.whiten1 = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(5.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+        self.whiten2 = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(5.0, ratio=3.0),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+    ) -> Tensor:
+        """.
+                Args:
+                   x: a Tensor of shape (seq_len, batch_size, num_channels)
+        attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len)
+                Returns:
+                   a Tensor with the same shape as x
+        """
+        x = self.in_proj(x)
+
+        (seq_len, batch_size, _) = x.shape
+        hidden_channels = self.hidden_channels
+
+        s, x, y = x.chunk(3, dim=-1)
+
+        # s will go through tanh.
+
+        s = self.balancer(s)
+        s = self.tanh(s)
+
+        s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels)
+        x = self.whiten1(x)
+        x = x * s
+        x = self.identity1(x)  # diagnostics only, it's the identity.
+
+        (seq_len, batch_size, embed_dim) = x.shape
+        num_heads = attn_weights.shape[0]
+        assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len)
+
+        x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3)
+        # now x: (num_heads, batch_size, seq_len, head_dim)
+        x = torch.matmul(attn_weights, x)
+        # now x: (num_heads, batch_size, seq_len, head_dim)
+        x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1)
+
+        y = self.identity2(y)
+        x = x * y
+        x = self.identity3(x)
+
+        x = self.out_proj(x)
+        x = self.whiten2(x)
+        return x
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        attn_weights: Tensor,
+        cached_x: Tensor,
+        left_context_len: int,
+    ) -> Tuple[Tensor, Tensor]:
+        """.
+        Args:
+            x: a Tensor of shape (seq_len, batch_size, num_channels)
+            attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len)
+            cached_x: left context, a Tensor of shape
+              (num_heads, batch_size, left_context_len, head_dim)
+            left_context_len: number of left context frames.
+        Returns:
+            - a Tensor with the same shape as x
+            - updated left context with same shape as cached_x
+        """
+        x = self.in_proj(x)
+
+        (seq_len, batch_size, _) = x.shape
+        hidden_channels = self.hidden_channels
+
+        s, x, y = x.chunk(3, dim=-1)
+
+        # s will go through tanh.
+        s = self.tanh(s)
+
+        s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels)
+        x = x * s
+
+        (seq_len, batch_size, embed_dim) = x.shape
+        num_heads = attn_weights.shape[0]
+        assert attn_weights.shape == (
+            num_heads,
+            batch_size,
+            seq_len,
+            left_context_len + seq_len,
+        )
+
+        x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3)
+        # now x: (num_heads, batch_size, seq_len, head_dim)
+
+        # Pad cached tensor
+        assert cached_x.shape[2] == left_context_len, (
+            cached_x.shape[2],
+            left_context_len,
+        )
+        x_pad = torch.cat([cached_x, x], dim=2)
+        # Update cached tensor
+        cached_x = x_pad[:, :, -left_context_len:, :]
+
+        x = torch.matmul(attn_weights, x_pad)
+        # now x: (num_heads, batch_size, seq_len, head_dim)
+        x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1)
+
+        x = x * y
+
+        x = self.out_proj(x)
+        return x, cached_x
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Zipformer2 model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/zipformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        causal: bool,
+    ) -> None:
+        """Construct a ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        bottleneck_dim = channels
+        self.causal = causal
+
+        self.in_proj = nn.Linear(
+            channels,
+            2 * bottleneck_dim,
+        )
+        # the gradients on in_proj are a little noisy, likely to do with the
+        # sigmoid in glu.
+
+        # after in_proj we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in
+        # the range 1 to 4, but sometimes, for some reason, for layer 0 the
+        # rms ends up being very large, between 50 and 100 for different channels.
+        # This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.balancer1 = Balancer(
+            bottleneck_dim,
+            channel_dim=-1,
+            min_positive=ScheduledFloat((0.0, 0.05), (8000.0, 0.025)),
+            max_positive=1.0,
+            min_abs=1.5,
+            max_abs=ScheduledFloat((0.0, 5.0), (8000.0, 10.0), default=1.0),
+        )
+
+        self.activation1 = Identity()  # for diagnostics
+
+        self.sigmoid = nn.Sigmoid()
+
+        self.activation2 = Identity()  # for diagnostics
+
+        assert kernel_size % 2 == 1
+
+        self.depthwise_conv = (
+            ChunkCausalDepthwiseConv1d(channels=bottleneck_dim, kernel_size=kernel_size)
+            if causal
+            else nn.Conv1d(
+                in_channels=bottleneck_dim,
+                out_channels=bottleneck_dim,
+                groups=bottleneck_dim,
+                kernel_size=kernel_size,
+                padding=kernel_size // 2,
+            )
+        )
+
+        self.balancer2 = Balancer(
+            bottleneck_dim,
+            channel_dim=1,
+            min_positive=ScheduledFloat((0.0, 0.1), (8000.0, 0.05)),
+            max_positive=1.0,
+            min_abs=ScheduledFloat((0.0, 0.2), (20000.0, 0.5)),
+            max_abs=10.0,
+        )
+
+        self.whiten = Whiten(
+            num_groups=1,
+            whitening_limit=_whitening_schedule(7.5),
+            prob=(0.025, 0.25),
+            grad_scale=0.01,
+        )
+
+        self.out_proj = ActivationDropoutAndLinear(
+            bottleneck_dim,
+            channels,
+            activation="SwooshR",
+            dropout_p=0.0,
+            initial_scale=0.05,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+        chunk_size: int = -1,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+           src_key_padding_mask: the mask for the src keys per batch (optional):
+               (batch, #time), contains True in masked positions.
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+
+        x = self.in_proj(x)  # (time, batch, 2*channels)
+
+        x, s = x.chunk(2, dim=-1)
+        s = self.balancer1(s)
+        s = self.sigmoid(s)
+        x = self.activation1(x)  # identity.
+        x = x * s
+        x = self.activation2(x)  # identity
+
+        # (time, batch, channels)
+
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        if src_key_padding_mask is not None:
+            x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+
+        if (
+            not torch.jit.is_scripting()
+            and not torch.jit.is_tracing()
+            and chunk_size >= 0
+        ):
+            # Not support exporting a model for simulated streaming decoding
+            assert (
+                self.causal
+            ), "Must initialize model with causal=True if you use chunk_size"
+            x = self.depthwise_conv(x, chunk_size=chunk_size)
+        else:
+            x = self.depthwise_conv(x)
+
+        x = self.balancer2(x)
+        x = x.permute(2, 0, 1)  # (time, batch, channels)
+
+        x = self.whiten(x)  # (time, batch, channels)
+        x = self.out_proj(x)  # (time, batch, channels)
+
+        return x
+
+    def streaming_forward(
+        self,
+        x: Tensor,
+        cache: Tensor,
+        src_key_padding_mask: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """Compute convolution module in streaming forward mode.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            cache: cached left context for depthwise_conv of shape
+              (#batch, channels, left_pad)
+            src_key_padding_mask: the mask for the src keys per batch (optional):
+              (batch, #time), contains True in masked positions.
+
+        Returns:
+            - Output tensor (#time, batch, channels).
+            - Updated cache (#batch, channels, left_pad)
+        """
+
+        x = self.in_proj(x)  # (time, batch, 2*channels)
+
+        x, s = x.chunk(2, dim=2)
+        s = self.sigmoid(s)
+        x = x * s
+        # (time, batch, channels)
+
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        if src_key_padding_mask is not None:
+            x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+
+        x, cache = self.depthwise_conv.streaming_forward(x, cache=cache)
+
+        x = x.permute(2, 0, 1)  # (time, batch, channels)
+
+        x = self.out_proj(x)  # (time, batch, channels)
+
+        return x, cache
+
+
+class ScalarMultiply(nn.Module):
+    def __init__(self, scale: float):
+        super().__init__()
+        self.scale = scale
+
+    def forward(self, x):
+        return x * self.scale
+
+
+def _test_zipformer_main(causal: bool = False):
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    memory_dim = 100
+
+    c = Zipformer2(
+        encoder_dim=(64, 96),
+        encoder_unmasked_dim=(48, 64),
+        num_heads=(4, 4),
+        causal=causal,
+        chunk_size=(4,) if causal else (-1,),
+        left_context_frames=(64,),
+        memory_dim=memory_dim,
+    )
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(seq_len, batch_size, 64),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+        memory=torch.randn(101, batch_size, memory_dim),
+    )
+    f[0].sum().backward()
+    c.eval()
+    f = c(
+        torch.randn(seq_len, batch_size, 64),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+    )
+    f  # to remove flake8 warnings
+
+
+if __name__ == "__main__":
+    logging.getLogger().setLevel(logging.INFO)
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    _test_zipformer_main(False)
+    _test_zipformer_main(True)
diff --git a/egs/librispeech/ASR/RESULTS.md b/egs/librispeech/ASR/RESULTS.md
index b945f43fd..fc7fcdc26 100644
--- a/egs/librispeech/ASR/RESULTS.md
+++ b/egs/librispeech/ASR/RESULTS.md
@@ -75,7 +75,7 @@ See <https://github.com/k2-fsa/icefall/pull/1058> for more details.
 ##### normal-scaled model, number of model parameters: 65549011, i.e., 65.55 M
 
 The tensorboard log can be found at
-<https://tensorboard.dev/experiment/cBaoIabCQxSDsyZM7FzqZA/>
+<https://tensorboard.dev/experiment/R2DT9Ju4QiadC4e2ioKh5A/>
 
 You can find a pretrained model, training logs, decoding logs, and decoding results at:
 <https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15>
@@ -90,18 +90,20 @@ You can use <https://github.com/k2-fsa/sherpa> to deploy it.
 | greedy_search        | 2.23       | 4.96       | --epoch 40 --avg 16 |
 | modified_beam_search | 2.21       | 4.91       | --epoch 40 --avg 16 |
 | fast_beam_search     | 2.24       | 4.93       | --epoch 40 --avg 16 |
+| greedy_search        | 2.22       | 4.87       | --epoch 50 --avg 25 |
+| modified_beam_search | 2.21       | 4.79       | --epoch 50 --avg 25 |
+| fast_beam_search     | 2.21       | 4.82       | --epoch 50 --avg 25 |
 | modified_beam_search_shallow_fusion | 2.01 | 4.37 | --epoch 40 --avg 16 --beam-size 12 --lm-scale 0.3 |
 | modified_beam_search_LODR | 1.94 | 4.17 | --epoch 40 --avg 16 --beam-size 12 --lm-scale 0.52 --LODR-scale -0.26 |
 | modified_beam_search_rescore | 2.04 | 4.39 | --epoch 40 --avg 16 --beam-size 12 |
 | modified_beam_search_rescore_LODR | 2.01  | 4.33 | --epoch 40 --avg 16 --beam-size 12  |
 
-
 The training command is:
 ```bash
 export CUDA_VISIBLE_DEVICES="0,1,2,3"
 ./zipformer/train.py \
   --world-size 4 \
-  --num-epochs 40 \
+  --num-epochs 50 \
   --start-epoch 1 \
   --use-fp16 1 \
   --exp-dir zipformer/exp \
@@ -115,8 +117,8 @@ The decoding command is:
 export CUDA_VISIBLE_DEVICES="0"
 for m in greedy_search modified_beam_search fast_beam_search; do
   ./zipformer/decode.py \
-    --epoch 30 \
-    --avg 9 \
+    --epoch 50 \
+    --avg 25 \
     --use-averaged-model 1 \
     --exp-dir ./zipformer/exp \
     --max-duration 600 \
@@ -129,7 +131,7 @@ To decode with external language models, please refer to the documentation [here
 ##### small-scaled model, number of model parameters: 23285615, i.e., 23.3 M
 
 The tensorboard log can be found at
-<https://tensorboard.dev/experiment/53P4tL22TpO0UdiL0kPaLg/>
+<https://tensorboard.dev/experiment/M9C8cYPWSN2MVBYaBIX3EQ/>
 
 You can find a pretrained model, training logs, decoding logs, and decoding results at:
 <https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-small-2023-05-16>
@@ -144,13 +146,16 @@ You can use <https://github.com/k2-fsa/sherpa> to deploy it.
 | greedy_search        | 2.49       | 5.91       | --epoch 40 --avg 13 |
 | modified_beam_search | 2.46       | 5.83       | --epoch 40 --avg 13 |
 | fast_beam_search     | 2.46       | 5.87       | --epoch 40 --avg 13 |
+| greedy_search        | 2.46       | 5.86       | --epoch 50 --avg 23 |
+| modified_beam_search | 2.42       | 5.73       | --epoch 50 --avg 23 |
+| fast_beam_search     | 2.46       | 5.78       | --epoch 50 --avg 23 |
 
 The training command is:
 ```bash
 export CUDA_VISIBLE_DEVICES="0,1"
 ./zipformer/train.py \
   --world-size 2 \
-  --num-epochs 40 \
+  --num-epochs 50 \
   --start-epoch 1 \
   --use-fp16 1 \
   --exp-dir zipformer/exp-small \
@@ -169,8 +174,8 @@ The decoding command is:
 export CUDA_VISIBLE_DEVICES="0"
 for m in greedy_search modified_beam_search fast_beam_search; do
   ./zipformer/decode.py \
-    --epoch 40 \
-    --avg 13 \
+    --epoch 50 \
+    --avg 23 \
     --exp-dir zipformer/exp-small \
     --max-duration 600 \
     --causal 0 \
@@ -185,7 +190,7 @@ done
 ##### large-scaled model, number of model parameters: 148439574, i.e., 148.4 M
 
 The tensorboard log can be found at
-<https://tensorboard.dev/experiment/HJ74wWYpQAGSzETkmQnrmQ/>
+<https://tensorboard.dev/experiment/C5ZPE5u1So2ZwhYLKW0FVg/>
 
 You can find a pretrained model, training logs, decoding logs, and decoding results at:
 <https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-large-2023-05-16>
@@ -200,13 +205,16 @@ You can use <https://github.com/k2-fsa/sherpa> to deploy it.
 | greedy_search        | 2.12       | 4.8        | --epoch 40 --avg 13 |
 | modified_beam_search | 2.11       | 4.7        | --epoch 40 --avg 13 |
 | fast_beam_search     | 2.13       | 4.78       | --epoch 40 --avg 13 |
+| greedy_search        | 2.08       | 4.69       | --epoch 50 --avg 30 |
+| modified_beam_search | 2.06       | 4.63       | --epoch 50 --avg 30 |
+| fast_beam_search     | 2.09       | 4.68       | --epoch 50 --avg 30 |
 
 The training command is:
 ```bash
 export CUDA_VISIBLE_DEVICES="0,1,2,3"
 ./zipformer/train.py \
   --world-size 4 \
-  --num-epochs 40 \
+  --num-epochs 50 \
   --start-epoch 1 \
   --use-fp16 1 \
   --exp-dir zipformer/exp-large \
@@ -224,8 +232,8 @@ The decoding command is:
 export CUDA_VISIBLE_DEVICES="0"
 for m in greedy_search modified_beam_search fast_beam_search; do
   ./zipformer/decode.py \
-    --epoch 40 \
-    --avg 16 \
+    --epoch 50 \
+    --avg 30 \
     --exp-dir zipformer/exp-large \
     --max-duration 600 \
     --causal 0 \
diff --git a/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_H.py b/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_H.py
new file mode 100755
index 000000000..4bdec9e11
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_H.py
@@ -0,0 +1,247 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This file shows how to use a torchscript model for decoding with H
+on CPU using OpenFST and decoders from kaldi.
+
+Usage:
+
+(1) LibriSpeech conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_H.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --H ./data/lang_bpe_500/H.fst \
+      --tokens ./data/lang_bpe_500/tokens.txt \
+      ./download/LibriSpeech/test-clean/1089/134686/1089-134686-0002.flac \
+      ./download/LibriSpeech/test-clean/1221/135766/1221-135766-0001.flac
+
+
+(2) AIShell conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_H.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --H ./data/lang_char/H.fst \
+      --tokens ./data/lang_char/tokens.txt \
+      ./BAC009S0764W0121.wav \
+      ./BAC009S0764W0122.wav \
+      ./BAC009S0764W0123.wav
+
+Note that to generate ./conformer_ctc/exp/cpu_jit.pt,
+you can use ./export.py --jit 1
+"""
+
+import argparse
+import logging
+import math
+from typing import Dict, List
+
+import kaldifeat
+import kaldifst
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="""Path to the torchscript model.
+        You can use ./conformer_ctc/export.py --jit 1
+        to obtain it
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to tokens.txt",
+    )
+
+    parser.add_argument("--H", type=str, required=True, help="Path to H.fst")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def read_tokens(tokens_txt: str) -> Dict[int, str]:
+    id2token = dict()
+    with open(tokens_txt, encoding="utf-8") as f:
+        for line in f:
+            token, idx = line.strip().split()
+            id2token[int(idx)] = token
+
+    return id2token
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    nnet_output: torch.Tensor,
+    H: kaldifst,
+    id2token: Dict[int, str],
+) -> List[str]:
+    """
+    Args:
+      filename:
+        Path to the filename for decoding. Used for debugging.
+      nnet_output:
+        A 2-D float32 tensor of shape (num_frames, vocab_size). It
+        contains output from log_softmax.
+      H:
+        The H graph.
+      id2token:
+        A map mapping token ID to token string.
+    Returns:
+      Return a list of decoded tokens.
+    """
+    logging.info(f"{filename}, {nnet_output.shape}")
+    decodable = DecodableCtc(nnet_output.cpu())
+
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(H, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        logging.info(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        logging.info(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # tokens are incremented during graph construction
+    # so they need to be decremented
+    hyps = [id2token[i - 1] for i in osymbols_out]
+    #  hyps = "".join(hyps).split("▁")
+    hyps = "".join(hyps).split("\u2581")  # unicode codepoint of ▁
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    logging.info("Loading torchscript model")
+    model = torch.jit.load(args.nn_model)
+    model.eval()
+    model.to(device)
+
+    logging.info(f"Loading H from {args.H}")
+    H = kaldifst.StdVectorFst.read(args.H)
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files, expected_sample_rate=sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.shape[0] for f in features]
+    feature_lengths = torch.tensor(feature_lengths)
+
+    supervisions = dict()
+    supervisions["sequence_idx"] = torch.arange(len(features))
+    supervisions["start_frame"] = torch.zeros(len(features))
+    supervisions["num_frames"] = feature_lengths
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    nnet_output, _, _ = model(features, supervisions)
+    feature_lengths = ((feature_lengths - 1) // 2 - 1) // 2
+
+    id2token = read_tokens(args.tokens)
+
+    hyps = []
+    for i in range(nnet_output.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[i],
+            nnet_output=nnet_output[i, : feature_lengths[i]],
+            H=H,
+            id2token=id2token,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py b/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py
new file mode 100755
index 000000000..d5a1dba3c
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py
@@ -0,0 +1,244 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This file shows how to use a torchscript model for decoding with HL
+on CPU using OpenFST and decoders from kaldi.
+
+Usage:
+
+(1) LibriSpeech conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_HL.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --HL ./data/lang_bpe_500/HL.fst \
+      --words ./data/lang_bpe_500/words.txt \
+      ./download/LibriSpeech/test-clean/1089/134686/1089-134686-0002.flac \
+      ./download/LibriSpeech/test-clean/1221/135766/1221-135766-0001.flac
+
+(2) AIShell conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_HL.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --HL ./data/lang_char/HL.fst \
+      --words ./data/lang_char/words.txt \
+      ./BAC009S0764W0121.wav \
+      ./BAC009S0764W0122.wav \
+      ./BAC009S0764W0123.wav
+
+
+Note that to generate ./conformer_ctc/exp/cpu_jit.pt,
+you can use ./export.py --jit 1
+"""
+
+import argparse
+import logging
+import math
+from typing import Dict, List
+
+import kaldifeat
+import kaldifst
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="""Path to the torchscript model.
+        You can use ./conformer_ctc/export.py --jit 1
+        to obtain it
+        """,
+    )
+
+    parser.add_argument(
+        "--words",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HL", type=str, required=True, help="Path to HL.fst")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def read_words(words_txt: str) -> Dict[int, str]:
+    id2word = dict()
+    with open(words_txt, encoding="utf-8") as f:
+        for line in f:
+            word, idx = line.strip().split()
+            id2word[int(idx)] = word
+
+    return id2word
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    nnet_output: torch.Tensor,
+    HL: kaldifst,
+    id2word: Dict[int, str],
+) -> List[str]:
+    """
+    Args:
+      filename:
+        Path to the filename for decoding. Used for debugging.
+      nnet_output:
+        A 2-D float32 tensor of shape (num_frames, vocab_size). It
+        contains output from log_softmax.
+      HL:
+        The HL graph.
+      id2word:
+        A map mapping word ID to word string.
+    Returns:
+      Return a list of decoded words.
+    """
+    logging.info(f"{filename}, {nnet_output.shape}")
+    decodable = DecodableCtc(nnet_output.cpu())
+
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(HL, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        logging.info(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        logging.info(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # are shifted by 1 during graph construction
+    hyps = [id2word[i] for i in osymbols_out]
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    logging.info("Loading torchscript model")
+    model = torch.jit.load(args.nn_model)
+    model.eval()
+    model.to(device)
+
+    logging.info(f"Loading HL from {args.HL}")
+    HL = kaldifst.StdVectorFst.read(args.HL)
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files, expected_sample_rate=sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.shape[0] for f in features]
+    feature_lengths = torch.tensor(feature_lengths)
+
+    supervisions = dict()
+    supervisions["sequence_idx"] = torch.arange(len(features))
+    supervisions["start_frame"] = torch.zeros(len(features))
+    supervisions["num_frames"] = feature_lengths
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    nnet_output, _, _ = model(features, supervisions)
+    feature_lengths = ((feature_lengths - 1) // 2 - 1) // 2
+
+    id2word = read_words(args.words)
+
+    hyps = []
+    for i in range(nnet_output.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[i],
+            nnet_output=nnet_output[i, : feature_lengths[i]],
+            HL=HL,
+            id2word=id2word,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py b/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py
new file mode 100755
index 000000000..216677a23
--- /dev/null
+++ b/egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py
@@ -0,0 +1,243 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This file shows how to use a torchscript model for decoding with HLG
+on CPU using OpenFST and decoders from kaldi.
+
+Usage:
+
+(1) LibriSpeech conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_HLG.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --HLG ./data/lang_bpe_500/HLG.fst \
+      --words ./data/lang_bpe_500/words.txt \
+      ./download/LibriSpeech/test-clean/1089/134686/1089-134686-0002.flac \
+      ./download/LibriSpeech/test-clean/1221/135766/1221-135766-0001.flac
+
+(2) AIShell conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_HLG.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --HLG ./data/lang_char/HLG.fst \
+      --words ./data/lang_char/words.txt \
+      ./BAC009S0764W0121.wav \
+      ./BAC009S0764W0122.wav \
+      ./BAC009S0764W0123.wav
+
+Note that to generate ./conformer_ctc/exp/cpu_jit.pt,
+you can use ./export.py --jit 1
+"""
+
+import argparse
+import logging
+import math
+from typing import Dict, List
+
+import kaldifeat
+import kaldifst
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="""Path to the torchscript model.
+        You can use ./conformer_ctc/export.py --jit 1
+        to obtain it
+        """,
+    )
+
+    parser.add_argument(
+        "--words",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HLG", type=str, required=True, help="Path to HLG.fst")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def read_words(words_txt: str) -> Dict[int, str]:
+    id2word = dict()
+    with open(words_txt, encoding="utf-8") as f:
+        for line in f:
+            word, idx = line.strip().split()
+            id2word[int(idx)] = word
+
+    return id2word
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    nnet_output: torch.Tensor,
+    HLG: kaldifst,
+    id2word: Dict[int, str],
+) -> List[str]:
+    """
+    Args:
+      filename:
+        Path to the filename for decoding. Used for debugging.
+      nnet_output:
+        A 2-D float32 tensor of shape (num_frames, vocab_size). It
+        contains output from log_softmax.
+      HLG:
+        The HLG graph.
+      id2word:
+        A map mapping word ID to word string.
+    Returns:
+      Return a list of decoded words.
+    """
+    logging.info(f"{filename}, {nnet_output.shape}")
+    decodable = DecodableCtc(nnet_output.cpu())
+
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(HLG, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        logging.info(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        logging.info(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # are shifted by 1 during graph construction
+    hyps = [id2word[i] for i in osymbols_out]
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    logging.info("Loading torchscript model")
+    model = torch.jit.load(args.nn_model)
+    model.eval()
+    model.to(device)
+
+    logging.info(f"Loading HLG from {args.HLG}")
+    HLG = kaldifst.StdVectorFst.read(args.HLG)
+
+    sample_rate = 16000
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files, expected_sample_rate=sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.shape[0] for f in features]
+    feature_lengths = torch.tensor(feature_lengths)
+
+    supervisions = dict()
+    supervisions["sequence_idx"] = torch.arange(len(features))
+    supervisions["start_frame"] = torch.zeros(len(features))
+    supervisions["num_frames"] = feature_lengths
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    nnet_output, _, _ = model(features, supervisions)
+    feature_lengths = ((feature_lengths - 1) // 2 - 1) // 2
+
+    id2word = read_words(args.words)
+
+    hyps = []
+    for i in range(nnet_output.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[i],
+            nnet_output=nnet_output[i, : feature_lengths[i]],
+            HLG=HLG,
+            id2word=id2word,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/conformer_ctc/train.py b/egs/librispeech/ASR/conformer_ctc/train.py
index 99fe64793..828106f41 100755
--- a/egs/librispeech/ASR/conformer_ctc/train.py
+++ b/egs/librispeech/ASR/conformer_ctc/train.py
@@ -557,7 +557,6 @@ def train_one_epoch(
             )
 
         if batch_idx % params.log_interval == 0:
-
             if tb_writer is not None:
                 loss_info.write_summary(
                     tb_writer, "train/current_", params.batch_idx_train
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py
index c5a05d349..ca21bd6bf 100755
--- a/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py
@@ -953,7 +953,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py
index 6bb37b017..23ddb6bec 100755
--- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py
@@ -953,7 +953,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train2.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train2.py
index 36067510c..420dc1065 100755
--- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train2.py
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train2.py
@@ -955,7 +955,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/distillation_with_hubert.sh b/egs/librispeech/ASR/distillation_with_hubert.sh
index 6aaa0333b..a5b0b85af 100755
--- a/egs/librispeech/ASR/distillation_with_hubert.sh
+++ b/egs/librispeech/ASR/distillation_with_hubert.sh
@@ -56,6 +56,8 @@ use_extracted_codebook=True
 #   "hubert_xtralarge_ll60k" -> pretrained model without fintuing
 teacher_model_id=hubert_xtralarge_ll60k_finetune_ls960
 
+. shared/parse_options.sh || exit 1
+
 log() {
   # This function is from espnet
   local fname=${BASH_SOURCE[1]##*/}
diff --git a/egs/librispeech/ASR/local/download_lm.py b/egs/librispeech/ASR/local/download_lm.py
index da1648d06..5a36ff2a9 100755
--- a/egs/librispeech/ASR/local/download_lm.py
+++ b/egs/librispeech/ASR/local/download_lm.py
@@ -43,6 +43,7 @@ from pathlib import Path
 
 from tqdm.auto import tqdm
 
+
 # This function is copied from lhotse
 def tqdm_urlretrieve_hook(t):
     """Wraps tqdm instance.
diff --git a/egs/librispeech/ASR/local/prepare_lang_fst.py b/egs/librispeech/ASR/local/prepare_lang_fst.py
new file mode 100755
index 000000000..fb1e7f9c0
--- /dev/null
+++ b/egs/librispeech/ASR/local/prepare_lang_fst.py
@@ -0,0 +1,219 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)
+
+"""
+This script takes as input lang_dir containing lexicon_disambig.txt,
+tokens.txt, and words.txt and generates the following files:
+
+    - H.fst
+    - HL.fst
+    - HLG.fst
+
+Note that saved files are in OpenFst binary format.
+
+Usage:
+
+./local/prepare_lang_fst.py \
+  --lang-dir ./data/lang_phone \
+  --has-silence 1
+
+Or
+
+./local/prepare_lang_fst.py \
+  --lang-dir ./data/lang_bpe_500
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import kaldifst
+
+from icefall.ctc import (
+    Lexicon,
+    add_disambig_self_loops,
+    add_one,
+    build_standard_ctc_topo,
+    make_lexicon_fst_no_silence,
+    make_lexicon_fst_with_silence,
+)
+from icefall.utils import str2bool
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--has-silence",
+        type=str2bool,
+        default=False,
+        help="True if the lexicon has silence.",
+    )
+
+    parser.add_argument(
+        "--ngram-G",
+        type=str,
+        help="""If not empty, it is the filename of G used to build HLG.
+        For instance, --ngram-G=./data/lm/G_3_fst.txt
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def build_HL(
+    H: kaldifst.StdVectorFst,
+    L: kaldifst.StdVectorFst,
+    has_silence: bool,
+    lexicon: Lexicon,
+) -> kaldifst.StdVectorFst:
+    if has_silence:
+        # We also need to change the input labels of L
+        add_one(L, treat_ilabel_zero_specially=True, update_olabel=False)
+    else:
+        add_one(L, treat_ilabel_zero_specially=False, update_olabel=False)
+
+    # Invoke add_disambig_self_loops() so that it eats the disambig symbols
+    # from L after composition
+    add_disambig_self_loops(
+        H,
+        start=lexicon.token2id["#0"] + 1,
+        end=lexicon.max_disambig_id + 1,
+    )
+
+    kaldifst.arcsort(H, sort_type="olabel")
+    kaldifst.arcsort(L, sort_type="ilabel")
+
+    HL = kaldifst.compose(H, L)
+    kaldifst.determinize_star(HL)
+
+    disambig0 = lexicon.token2id["#0"] + 1
+    max_disambig = lexicon.max_disambig_id + 1
+    for state in kaldifst.StateIterator(HL):
+        for arc in kaldifst.ArcIterator(HL, state):
+            # If treat_ilabel_zero_specially is False, we always change it
+            # Otherwise, we only change non-zero input labels
+            if disambig0 <= arc.ilabel <= max_disambig:
+                arc.ilabel = 0
+
+    # Note: We are not composing L with G, so there is no need to add
+    # self-loops to L to handle #0
+
+    return HL
+
+
+def build_HLG(
+    H: kaldifst.StdVectorFst,
+    L: kaldifst.StdVectorFst,
+    G: kaldifst.StdVectorFst,
+    has_silence: bool,
+    lexicon: Lexicon,
+) -> kaldifst.StdVectorFst:
+    if has_silence:
+        # We also need to change the input labels of L
+        add_one(L, treat_ilabel_zero_specially=True, update_olabel=False)
+    else:
+        add_one(L, treat_ilabel_zero_specially=False, update_olabel=False)
+
+    # add-self-loops
+    token_disambig0 = lexicon.token2id["#0"] + 1
+    word_disambig0 = lexicon.word2id["#0"]
+
+    kaldifst.add_self_loops(L, isyms=[token_disambig0], osyms=[word_disambig0])
+
+    kaldifst.arcsort(L, sort_type="olabel")
+    kaldifst.arcsort(G, sort_type="ilabel")
+    LG = kaldifst.compose(L, G)
+    kaldifst.determinize_star(LG)
+    kaldifst.minimize_encoded(LG)
+
+    kaldifst.arcsort(LG, sort_type="ilabel")
+
+    # Invoke add_disambig_self_loops() so that it eats the disambig symbols
+    # from L after composition
+    add_disambig_self_loops(
+        H,
+        start=lexicon.token2id["#0"] + 1,
+        end=lexicon.max_disambig_id + 1,
+    )
+
+    kaldifst.arcsort(H, sort_type="olabel")
+
+    HLG = kaldifst.compose(H, LG)
+    kaldifst.determinize_star(HLG)
+
+    disambig0 = lexicon.token2id["#0"] + 1
+    max_disambig = lexicon.max_disambig_id + 1
+    for state in kaldifst.StateIterator(HLG):
+        for arc in kaldifst.ArcIterator(HLG, state):
+            # If treat_ilabel_zero_specially is False, we always change it
+            # Otherwise, we only change non-zero input labels
+            if disambig0 <= arc.ilabel <= max_disambig:
+                arc.ilabel = 0
+    return HLG
+
+
+def copy_fst(fst):
+    # Please don't use fst.copy()
+    return kaldifst.StdVectorFst(fst)
+
+
+def main():
+    args = get_args()
+    lang_dir = args.lang_dir
+
+    lexicon = Lexicon(lang_dir)
+
+    logging.info("Building standard CTC topology")
+    max_token_id = max(lexicon.tokens)
+    H = build_standard_ctc_topo(max_token_id=max_token_id)
+
+    # We need to add one to all tokens since we want to use ID 0
+    # for epsilon
+    add_one(H, treat_ilabel_zero_specially=False, update_olabel=True)
+    H.write(f"{lang_dir}/H.fst")
+
+    logging.info("Building L")
+    # Now for HL
+
+    if args.has_silence:
+        L = make_lexicon_fst_with_silence(lexicon, attach_symbol_table=False)
+    else:
+        L = make_lexicon_fst_no_silence(lexicon, attach_symbol_table=False)
+
+    logging.info("Building HL")
+    HL = build_HL(
+        H=copy_fst(H),
+        L=copy_fst(L),
+        has_silence=args.has_silence,
+        lexicon=lexicon,
+    )
+    HL.write(f"{lang_dir}/HL.fst")
+
+    if not args.ngram_G:
+        logging.info("Skip building HLG")
+        return
+
+    logging.info("Building HLG")
+    with open(args.ngram_G) as f:
+        G = kaldifst.compile(
+            s=f.read(),
+            acceptor=False,
+        )
+
+    HLG = build_HLG(H=H, L=L, G=G, has_silence=args.has_silence, lexicon=lexicon)
+    HLG.write(f"{lang_dir}/HLG.fst")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/long_file_recog/beam_search.py b/egs/librispeech/ASR/long_file_recog/beam_search.py
index f8c31861c..b65e9d40a 100644
--- a/egs/librispeech/ASR/long_file_recog/beam_search.py
+++ b/egs/librispeech/ASR/long_file_recog/beam_search.py
@@ -236,7 +236,7 @@ def greedy_search_batch(
     encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
 
     offset = 0
-    for (t, batch_size) in enumerate(batch_size_list):
+    for t, batch_size in enumerate(batch_size_list):
         start = offset
         end = offset + batch_size
         current_encoder_out = encoder_out.data[start:end]
@@ -507,7 +507,7 @@ def modified_beam_search(
 
     offset = 0
     finalized_B = []
-    for (t, batch_size) in enumerate(batch_size_list):
+    for t, batch_size in enumerate(batch_size_list):
         start = offset
         end = offset + batch_size
         current_encoder_out = encoder_out.data[start:end]
diff --git a/egs/librispeech/ASR/long_file_recog/merge_chunks.py b/egs/librispeech/ASR/long_file_recog/merge_chunks.py
index d38d9c86a..9e31e00d5 100755
--- a/egs/librispeech/ASR/long_file_recog/merge_chunks.py
+++ b/egs/librispeech/ASR/long_file_recog/merge_chunks.py
@@ -162,7 +162,6 @@ def merge_chunks(
 
     futures = []
     with ThreadPoolExecutor(max_workers=1) as executor:
-
         for cut in cuts_chunk:
             cur_rec_id = cut.recording.id
             if len(cut_list) == 0:
diff --git a/egs/librispeech/ASR/long_file_recog/recognize.py b/egs/librispeech/ASR/long_file_recog/recognize.py
index 96c83f859..466253446 100755
--- a/egs/librispeech/ASR/long_file_recog/recognize.py
+++ b/egs/librispeech/ASR/long_file_recog/recognize.py
@@ -264,6 +264,7 @@ def decode_dataset(
         - timestamps of reference transcript
         - timestamps of predicted result
     """
+
     #  Background worker to add alignemnt and save cuts to disk.
     def _save_worker(
         cuts: List[Cut],
diff --git a/egs/librispeech/ASR/lstm_transducer_stateless/test_model.py b/egs/librispeech/ASR/lstm_transducer_stateless/test_model.py
index 03dfe1997..91ef53e24 100755
--- a/egs/librispeech/ASR/lstm_transducer_stateless/test_model.py
+++ b/egs/librispeech/ASR/lstm_transducer_stateless/test_model.py
@@ -57,8 +57,7 @@ def test_model():
 
     convert_scaled_to_non_scaled(model, inplace=True)
 
-    if not os.path.exists(params.exp_dir):
-        os.path.mkdir(params.exp_dir)
+    params.exp_dir.mkdir(exist_ok=True)
 
     encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
     export_encoder_model_jit_trace(model.encoder, encoder_filename)
diff --git a/egs/librispeech/ASR/prepare.sh b/egs/librispeech/ASR/prepare.sh
index 8ce1eb478..739608572 100755
--- a/egs/librispeech/ASR/prepare.sh
+++ b/egs/librispeech/ASR/prepare.sh
@@ -299,6 +299,16 @@ if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
       --max-order=4 \
       $dl_dir/lm/4-gram.arpa > data/lm/G_4_gram.fst.txt
   fi
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+
+    if [ ! -f $lang_dir/HL.fst ]; then
+      ./local/prepare_lang_fst.py  \
+        --lang-dir $lang_dir \
+        --ngram-G ./data/lm/G_3_gram.fst.txt
+    fi
+  done
 fi
 
 if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
diff --git a/egs/librispeech/ASR/pruned2_knowledge/asr_datamodule.py b/egs/librispeech/ASR/pruned2_knowledge/asr_datamodule.py
index f8f558ce1..ee7556e49 100644
--- a/egs/librispeech/ASR/pruned2_knowledge/asr_datamodule.py
+++ b/egs/librispeech/ASR/pruned2_knowledge/asr_datamodule.py
@@ -209,7 +209,7 @@ class LibriSpeechAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "cuts_musan.json.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/librispeech/ASR/pruned2_knowledge/optim.py b/egs/librispeech/ASR/pruned2_knowledge/optim.py
index 76cd4e11e..9f287ce70 100644
--- a/egs/librispeech/ASR/pruned2_knowledge/optim.py
+++ b/egs/librispeech/ASR/pruned2_knowledge/optim.py
@@ -66,7 +66,6 @@ class Eve(Optimizer):
         weight_decay=1e-3,
         target_rms=0.1,
     ):
-
         if not 0.0 <= lr:
             raise ValueError("Invalid learning rate: {}".format(lr))
         if not 0.0 <= eps:
diff --git a/egs/librispeech/ASR/pruned2_knowledge/train.py b/egs/librispeech/ASR/pruned2_knowledge/train.py
index 77e06d3b7..a4899f7bd 100755
--- a/egs/librispeech/ASR/pruned2_knowledge/train.py
+++ b/egs/librispeech/ASR/pruned2_knowledge/train.py
@@ -811,7 +811,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py
index 3298568a3..7fcd242fc 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py
@@ -719,7 +719,7 @@ def greedy_search_batch(
     encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
 
     offset = 0
-    for (t, batch_size) in enumerate(batch_size_list):
+    for t, batch_size in enumerate(batch_size_list):
         start = offset
         end = offset + batch_size
         current_encoder_out = encoder_out.data[start:end]
@@ -1019,7 +1019,7 @@ def modified_beam_search(
 
     offset = 0
     finalized_B = []
-    for (t, batch_size) in enumerate(batch_size_list):
+    for t, batch_size in enumerate(batch_size_list):
         start = offset
         end = offset + batch_size
         current_encoder_out = encoder_out.data[start:end]
@@ -1227,7 +1227,7 @@ def modified_beam_search_lm_rescore(
 
     offset = 0
     finalized_B = []
-    for (t, batch_size) in enumerate(batch_size_list):
+    for t, batch_size in enumerate(batch_size_list):
         start = offset
         end = offset + batch_size
         current_encoder_out = encoder_out.data[start:end]
@@ -1427,7 +1427,7 @@ def modified_beam_search_lm_rescore_LODR(
 
     offset = 0
     finalized_B = []
-    for (t, batch_size) in enumerate(batch_size_list):
+    for t, batch_size in enumerate(batch_size_list):
         start = offset
         end = offset + batch_size
         current_encoder_out = encoder_out.data[start:end]
@@ -2608,7 +2608,6 @@ def modified_beam_search_LODR(
                 context_score = 0
                 new_context_state = None if context_graph is None else hyp.context_state
                 if new_token not in (blank_id, unk_id):
-
                     if context_graph is not None:
                         (
                             context_score,
@@ -2758,7 +2757,7 @@ def modified_beam_search_lm_shallow_fusion(
 
     offset = 0
     finalized_B = []
-    for (t, batch_size) in enumerate(batch_size_list):
+    for t, batch_size in enumerate(batch_size_list):
         start = offset
         end = offset + batch_size
         current_encoder_out = encoder_out.data[start:end]  # get batch
@@ -2900,7 +2899,6 @@ def modified_beam_search_lm_shallow_fusion(
                 new_token = topk_token_indexes[k]
                 new_timestamp = hyp.timestamp[:]
                 if new_token not in (blank_id, unk_id):
-
                     ys.append(new_token)
                     new_timestamp.append(t)
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py
index bcd419fb7..ab46e233b 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py
@@ -158,7 +158,7 @@ class Conformer(EncoderInterface):
         if not is_jit_tracing():
             assert x.size(0) == lengths.max().item()
 
-        src_key_padding_mask = make_pad_mask(lengths)
+        src_key_padding_mask = make_pad_mask(lengths, x.size(0))
 
         if self.dynamic_chunk_training:
             assert (
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless2/optim.py
index 2d7f557ad..f54bc2709 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless2/optim.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/optim.py
@@ -66,7 +66,6 @@ class Eve(Optimizer):
         weight_decay=1e-3,
         target_rms=0.1,
     ):
-
         if not 0.0 <= lr:
             raise ValueError("Invalid learning rate: {}".format(lr))
         if not 0.0 <= eps:
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py
index 963ebdc2d..91d64c1df 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py
@@ -528,7 +528,6 @@ class ScaledLSTM(nn.LSTM):
             return
 
         with torch.cuda.device_of(first_fw):
-
             # Note: no_grad() is necessary since _cudnn_rnn_flatten_weight is
             # an inplace operation on self._flat_weights
             with torch.no_grad():
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py
index b7735be85..057624272 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py
@@ -164,7 +164,7 @@ class AsrDataModule:
         if cuts_musan is not None:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/train.py b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py
index 3b5a635e4..66dc5f991 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless5/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py
@@ -1003,7 +1003,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py b/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py
index 14ff86f23..3bca7db2c 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py
@@ -56,7 +56,6 @@ class CodebookIndexExtractor:
     """
 
     def __init__(self, params: AttributeDict):
-
         self.params = params
         params.subsets = ["clean-100"]
         if self.params.full_libri:
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/alignment.py b/egs/librispeech/ASR/pruned_transducer_stateless7/alignment.py
index 76cd56bbb..bfb5fe609 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless7/alignment.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/alignment.py
@@ -111,7 +111,7 @@ def batch_force_alignment(
 
     offset = 0
     finalized_B = []
-    for (t, batch_size) in enumerate(batch_size_list):
+    for t, batch_size in enumerate(batch_size_list):
         start = offset
         end = offset + batch_size
         current_encoder_out = encoder_out.data[start:end]
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/finetune.py b/egs/librispeech/ASR/pruned_transducer_stateless7/finetune.py
index 3ee2b7d65..a7a8ef149 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless7/finetune.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/finetune.py
@@ -655,8 +655,12 @@ def load_model_params(
         dst_state_dict = model.state_dict()
         for module in init_modules:
             logging.info(f"Loading parameters starting with prefix {module}")
-            src_keys = [k for k in src_state_dict.keys() if k.startswith(module)]
-            dst_keys = [k for k in dst_state_dict.keys() if k.startswith(module)]
+            src_keys = [
+                k for k in src_state_dict.keys() if k.startswith(module.strip() + ".")
+            ]
+            dst_keys = [
+                k for k in dst_state_dict.keys() if k.startswith(module.strip() + ".")
+            ]
             assert set(src_keys) == set(dst_keys)  # two sets should match exactly
             for key in src_keys:
                 dst_state_dict[key] = src_state_dict.pop(key)
@@ -1089,6 +1093,9 @@ def run(rank, world_size, args):
         checkpoints = load_model_params(
             ckpt=params.finetune_ckpt, model=model, init_modules=modules
         )
+        if rank == 0:
+            # model_avg is only used with rank 0
+            model_avg = copy.deepcopy(model).to(torch.float64)
     else:
         assert params.start_epoch > 0, params.start_epoch
         checkpoints = load_checkpoint_if_available(
@@ -1132,7 +1139,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech.py b/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech.py
index 75e153cb0..cd432fd6f 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech.py
@@ -217,7 +217,7 @@ class GigaSpeechAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/model.py b/egs/librispeech/ASR/pruned_transducer_stateless7/model.py
index 0e59b0f2f..add0e6a18 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless7/model.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/model.py
@@ -114,6 +114,9 @@ class Transducer(nn.Module):
 
         assert x.size(0) == x_lens.size(0) == y.dim0
 
+        # x.T_dim == max(x_len)
+        assert x.size(1) == x_lens.max().item(), (x.shape, x_lens, x_lens.max())
+
         encoder_out, x_lens = self.encoder(x, x_lens)
         assert torch.all(x_lens > 0)
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless7/optim.py
index aa3cef338..8ab3589da 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless7/optim.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/optim.py
@@ -117,7 +117,7 @@ class BatchedOptimizer(Optimizer):
 
         yield tuples  # <-- calling code will do the actual optimization here!
 
-        for ((stacked_params, _state, _names), batch) in zip(tuples, batches):
+        for (stacked_params, _state, _names), batch in zip(tuples, batches):
             for i, p in enumerate(batch):  # batch is list of Parameter
                 p.copy_(stacked_params[i])
 
@@ -181,7 +181,6 @@ class ScaledAdam(BatchedOptimizer):
         parameters_names=None,
         show_dominant_parameters=True,
     ):
-
         assert parameters_names is not None, (
             "Please prepare parameters_names,"
             "which is a List[List[str]]. Each List[str] is for a group"
@@ -224,9 +223,7 @@ class ScaledAdam(BatchedOptimizer):
         batch = True
 
         for group, group_params_names in zip(self.param_groups, self.parameters_names):
-
             with self.batched_params(group["params"], group_params_names) as batches:
-
                 # batches is list of pairs (stacked_param, state).  stacked_param is like
                 # a regular parameter, and will have a .grad, but the 1st dim corresponds to
                 # a stacking dim, it is not a real dim.
@@ -325,7 +322,7 @@ class ScaledAdam(BatchedOptimizer):
         clipping_update_period = group["clipping_update_period"]
 
         tot_sumsq = torch.tensor(0.0, device=first_p.device)
-        for (p, state, param_names) in tuples:
+        for p, state, param_names in tuples:
             grad = p.grad
             if grad.is_sparse:
                 raise RuntimeError(
@@ -410,7 +407,7 @@ class ScaledAdam(BatchedOptimizer):
                 from tuples, we still pass it to save some time.
         """
         all_sumsq_orig = {}
-        for (p, state, batch_param_names) in tuples:
+        for p, state, batch_param_names in tuples:
             # p is a stacked batch parameters.
             batch_grad = p.grad
             if p.numel() == p.shape[0]:  # a batch of scalars
@@ -426,7 +423,6 @@ class ScaledAdam(BatchedOptimizer):
             for name, sumsq_orig, rms, grad in zip(
                 batch_param_names, batch_sumsq_orig, batch_rms_orig, batch_grad
             ):
-
                 proportion_orig = sumsq_orig / tot_sumsq
                 all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad)
 
@@ -1039,7 +1035,7 @@ def _test_scaled_adam(hidden_dim: int):
 
             # if epoch == 130:
             #    opts = diagnostics.TensorDiagnosticOptions(
-            #        2 ** 22
+            #        512
             #    )  # allow 4 megabytes per sub-module
             #    diagnostic = diagnostics.attach_diagnostics(m, opts)
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py
index 2b4d51089..fac3706d2 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless7/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py
@@ -1028,7 +1028,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py
index 5b75b8d35..cbde2a2e4 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py
@@ -281,7 +281,7 @@ class Zipformer(EncoderInterface):
 
         lengths = (x_lens - 7) >> 1
         assert x.size(0) == lengths.max().item(), (x.shape, lengths, lengths.max())
-        mask = make_pad_mask(lengths)
+        mask = make_pad_mask(lengths, x.size(0))
 
         outputs = []
         feature_masks = self.get_feature_masks(x)
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py
index b387968a9..d8fa08372 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py
@@ -1052,7 +1052,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py
index 23fb6f497..25a1aa674 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py
@@ -1042,7 +1042,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py
index 99090b2c1..2d915ff87 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py
@@ -1029,7 +1029,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train2.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train2.py
index 9be629149..aa6c0668a 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train2.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train2.py
@@ -1030,7 +1030,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py
index b494253d6..565dc7a16 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py
@@ -1141,7 +1141,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/export-onnx.py b/egs/librispeech/ASR/pruned_transducer_stateless8/export-onnx.py
new file mode 100755
index 000000000..3fef231a1
--- /dev/null
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/export-onnx.py
@@ -0,0 +1,604 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/WeijiZhuang/icefall-asr-librispeech-pruned-transducer-stateless8-2022-12-02
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/WeijiZhuang/icefall-asr-librispeech-pruned-transducer-stateless8-2022-12-02
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless8/export-onnx.py \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --feedforward-dims "1024,1024,2048,2048,1024"
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+from zipformer import Zipformer
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import num_tokens, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Zipformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Zipformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Zipformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "zipformer",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "stateless7",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=encoder_filename, meta_data=meta_data)
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, joiner_dim)
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    decoder_model = torch.jit.script(decoder_model)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    logging.info(f"device: {device}")
+
+    # <blk> is defined in local/train_bpe_model.py
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params, enable_giga=False)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul", "Gather"],
+        weight_type=QuantType.QInt8,
+    )
+
+    joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx"
+    quantize_dynamic(
+        model_input=joiner_filename,
+        model_output=joiner_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/train.py b/egs/librispeech/ASR/pruned_transducer_stateless8/train.py
index bee414292..3f271c5b4 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless8/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless8/train.py
@@ -1154,7 +1154,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/streaming_conformer_ctc/conformer.py b/egs/librispeech/ASR/streaming_conformer_ctc/conformer.py
index be6fabf35..0b982f4bf 100644
--- a/egs/librispeech/ASR/streaming_conformer_ctc/conformer.py
+++ b/egs/librispeech/ASR/streaming_conformer_ctc/conformer.py
@@ -230,7 +230,9 @@ class Conformer(Transformer):
                 x, pos_emb, mask=mask, src_key_padding_mask=src_key_padding_mask
             )  # (T, B, F)
         else:
-            x = self.encoder(x, pos_emb, src_key_padding_mask=src_key_padding_mask)  # (T, B, F)
+            x = self.encoder(
+                x, pos_emb, src_key_padding_mask=src_key_padding_mask
+            )  # (T, B, F)
 
         if self.normalize_before:
             x = self.after_norm(x)
diff --git a/egs/librispeech/ASR/streaming_conformer_ctc/train.py b/egs/librispeech/ASR/streaming_conformer_ctc/train.py
index bb55ed6bb..14d7274c2 100755
--- a/egs/librispeech/ASR/streaming_conformer_ctc/train.py
+++ b/egs/librispeech/ASR/streaming_conformer_ctc/train.py
@@ -543,7 +543,6 @@ def train_one_epoch(
             )
 
         if batch_idx % params.log_interval == 0:
-
             if tb_writer is not None:
                 loss_info.write_summary(
                     tb_writer, "train/current_", params.batch_idx_train
diff --git a/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py b/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py
index 20df469da..c500eb3e5 100644
--- a/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py
+++ b/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py
@@ -233,7 +233,7 @@ class LibriSpeechAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/librispeech/ASR/tdnn_lstm_ctc/train.py b/egs/librispeech/ASR/tdnn_lstm_ctc/train.py
index 0aa1587ba..90245ed46 100755
--- a/egs/librispeech/ASR/tdnn_lstm_ctc/train.py
+++ b/egs/librispeech/ASR/tdnn_lstm_ctc/train.py
@@ -463,7 +463,6 @@ def train_one_epoch(
                 f"tot_loss[{tot_loss}], batch size: {batch_size}"
             )
         if batch_idx % params.log_interval == 0:
-
             if tb_writer is not None:
                 loss_info.write_summary(
                     tb_writer, "train/current_", params.batch_idx_train
diff --git a/egs/librispeech/ASR/transducer/train.py b/egs/librispeech/ASR/transducer/train.py
index f2a09346c..9ac6b7d03 100755
--- a/egs/librispeech/ASR/transducer/train.py
+++ b/egs/librispeech/ASR/transducer/train.py
@@ -513,7 +513,6 @@ def train_one_epoch(
             )
 
         if batch_idx % params.log_interval == 0:
-
             if tb_writer is not None:
                 loss_info.write_summary(
                     tb_writer, "train/current_", params.batch_idx_train
diff --git a/egs/librispeech/ASR/transducer_lstm/train.py b/egs/librispeech/ASR/transducer_lstm/train.py
index a6f2bd08c..92134116c 100755
--- a/egs/librispeech/ASR/transducer_lstm/train.py
+++ b/egs/librispeech/ASR/transducer_lstm/train.py
@@ -517,7 +517,6 @@ def train_one_epoch(
             )
 
         if batch_idx % params.log_interval == 0:
-
             if tb_writer is not None:
                 loss_info.write_summary(
                     tb_writer, "train/current_", params.batch_idx_train
diff --git a/egs/librispeech/ASR/zipformer/decoder.py b/egs/librispeech/ASR/zipformer/decoder.py
index e8db988f6..e77e54118 100644
--- a/egs/librispeech/ASR/zipformer/decoder.py
+++ b/egs/librispeech/ASR/zipformer/decoder.py
@@ -61,10 +61,15 @@ class Decoder(nn.Module):
         )
         # the balancers are to avoid any drift in the magnitude of the
         # embeddings, which would interact badly with parameter averaging.
-        self.balancer = Balancer(decoder_dim, channel_dim=-1,
-                                 min_positive=0.0, max_positive=1.0,
-                                 min_abs=0.5, max_abs=1.0,
-                                 prob=0.05)
+        self.balancer = Balancer(
+            decoder_dim,
+            channel_dim=-1,
+            min_positive=0.0,
+            max_positive=1.0,
+            min_abs=0.5,
+            max_abs=1.0,
+            prob=0.05,
+        )
 
         self.blank_id = blank_id
 
@@ -81,10 +86,15 @@ class Decoder(nn.Module):
                 groups=decoder_dim // 4,  # group size == 4
                 bias=False,
             )
-            self.balancer2 = Balancer(decoder_dim, channel_dim=-1,
-                                      min_positive=0.0, max_positive=1.0,
-                                      min_abs=0.5, max_abs=1.0,
-                                      prob=0.05)
+            self.balancer2 = Balancer(
+                decoder_dim,
+                channel_dim=-1,
+                min_positive=0.0,
+                max_positive=1.0,
+                min_abs=0.5,
+                max_abs=1.0,
+                prob=0.05,
+            )
 
     def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
         """
@@ -107,9 +117,7 @@ class Decoder(nn.Module):
         if self.context_size > 1:
             embedding_out = embedding_out.permute(0, 2, 1)
             if need_pad is True:
-                embedding_out = F.pad(
-                    embedding_out, pad=(self.context_size - 1, 0)
-                )
+                embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0))
             else:
                 # During inference time, there is no need to do extra padding
                 # as we only need one output
diff --git a/egs/librispeech/ASR/zipformer/export-onnx-ctc.py b/egs/librispeech/ASR/zipformer/export-onnx-ctc.py
new file mode 100755
index 000000000..3345d20d3
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/export-onnx-ctc.py
@@ -0,0 +1,436 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a CTC model from PyTorch to ONNX.
+
+Note that the model is trained using both transducer and CTC loss. This script
+exports only the CTC head.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+pushd $repo
+git lfs pull --include "exp/pretrained.pt"
+
+cd exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./zipformer/export-onnx-ctc.py \
+  --use-transducer 0 \
+  --use-ctc 1 \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers "2,2,3,4,3,2" \
+  --downsampling-factor "1,2,4,8,4,2" \
+  --feedforward-dim "512,768,1024,1536,1024,768" \
+  --num-heads "4,4,4,8,4,4" \
+  --encoder-dim "192,256,384,512,384,256" \
+  --query-head-dim 32 \
+  --value-head-dim 12 \
+  --pos-head-dim 4 \
+  --pos-dim 48 \
+  --encoder-unmasked-dim "192,192,256,256,256,192" \
+  --cnn-module-kernel "31,31,15,15,15,31" \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --causal False \
+  --chunk-size 16 \
+  --left-context-frames 128
+
+It will generate the following 2 files inside $repo/exp:
+
+  - model.onnx
+  - model.int8.onnx
+
+See ./onnx_pretrained_ctc.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import k2
+import onnx
+import torch
+import torch.nn as nn
+from decoder import Decoder
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_model, get_params
+from zipformer import Zipformer2
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import make_pad_mask, num_tokens, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="zipformer/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        default="data/lang_bpe_500/tokens.txt",
+        help="Path to the tokens.txt",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxModel(nn.Module):
+    """A wrapper for encoder_embed, Zipformer, and ctc_output layer"""
+
+    def __init__(
+        self,
+        encoder: Zipformer2,
+        encoder_embed: nn.Module,
+        ctc_output: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            A Zipformer encoder.
+          encoder_embed:
+            The first downsampling layer for zipformer.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_embed = encoder_embed
+        self.ctc_output = ctc_output
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Zipformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - log_probs, a 3-D tensor of shape (N, T', vocab_size)
+            - log_probs_len, a 1-D int64 tensor of shape (N,)
+        """
+        x, x_lens = self.encoder_embed(x, x_lens)
+        src_key_padding_mask = make_pad_mask(x_lens)
+        x = x.permute(1, 0, 2)
+        encoder_out, log_probs_len = self.encoder(x, x_lens, src_key_padding_mask)
+        encoder_out = encoder_out.permute(1, 0, 2)
+        log_probs = self.ctc_output(encoder_out)
+
+        return log_probs, log_probs_len
+
+
+def export_ctc_model_onnx(
+    model: OnnxModel,
+    filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - log_probs, a tensor of shape (N, T', joiner_dim)
+        - log_probs_len, a tensor of shape (N,)
+
+    Args:
+      model:
+        The input model
+      filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    model = torch.jit.trace(model, (x, x_lens))
+
+    torch.onnx.export(
+        model,
+        (x, x_lens),
+        filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["log_probs", "log_probs_len"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "log_probs": {0: "N", 1: "T"},
+            "log_probs_len": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "model_type": "zipformer2_ctc",
+        "version": "1",
+        "model_author": "k2-fsa",
+        "comment": "non-streaming zipformer2 CTC",
+    }
+    logging.info(f"meta_data: {meta_data}")
+
+    add_meta_data(filename=filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    token_table = k2.SymbolTable.from_file(params.tokens)
+    params.blank_id = token_table["<blk>"]
+    params.vocab_size = num_tokens(token_table) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_model(params)
+
+    model.to(device)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        elif params.avg == 1:
+            load_checkpoint(
+                f"{params.exp_dir}/epoch-{params.epoch}.pt", model, strict=False
+            )
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True, is_onnx=True)
+
+    model = OnnxModel(
+        encoder=model.encoder,
+        encoder_embed=model.encoder_embed,
+        ctc_output=model.ctc_output,
+    )
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"num parameters: {num_param}")
+
+    opset_version = 13
+
+    logging.info("Exporting ctc model")
+    filename = params.exp_dir / f"model.onnx"
+    export_ctc_model_onnx(
+        model,
+        filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported to {filename}")
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    logging.info("Generate int8 quantization models")
+
+    filename_int8 = params.exp_dir / f"model.int8.onnx"
+    quantize_dynamic(
+        model_input=filename,
+        model_output=filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/joiner.py b/egs/librispeech/ASR/zipformer/joiner.py
index f03cc930e..dfb0a0057 100644
--- a/egs/librispeech/ASR/zipformer/joiner.py
+++ b/egs/librispeech/ASR/zipformer/joiner.py
@@ -52,12 +52,13 @@ class Joiner(nn.Module):
         Returns:
           Return a tensor of shape (N, T, s_range, C).
         """
-        assert encoder_out.ndim == decoder_out.ndim, (encoder_out.shape, decoder_out.shape)
+        assert encoder_out.ndim == decoder_out.ndim, (
+            encoder_out.shape,
+            decoder_out.shape,
+        )
 
         if project_input:
-            logit = self.encoder_proj(encoder_out) + self.decoder_proj(
-                decoder_out
-            )
+            logit = self.encoder_proj(encoder_out) + self.decoder_proj(decoder_out)
         else:
             logit = encoder_out + decoder_out
 
diff --git a/egs/librispeech/ASR/zipformer/onnx_decode.py b/egs/librispeech/ASR/zipformer/onnx_decode.py
index 2aca36ca9..356c2a830 100755
--- a/egs/librispeech/ASR/zipformer/onnx_decode.py
+++ b/egs/librispeech/ASR/zipformer/onnx_decode.py
@@ -303,7 +303,9 @@ def main():
 
     for test_set, test_dl in zip(test_sets, test_dl):
         start_time = time.time()
-        results, total_duration = decode_dataset(dl=test_dl, model=model, token_table=token_table)
+        results, total_duration = decode_dataset(
+            dl=test_dl, model=model, token_table=token_table
+        )
         end_time = time.time()
         elapsed_seconds = end_time - start_time
         rtf = elapsed_seconds / total_duration
diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc.py b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc.py
new file mode 100755
index 000000000..eb5cee9cd
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc.py
@@ -0,0 +1,213 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+"""
+This script loads ONNX models and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+as an example to show how to use this file.
+
+1. Please follow ./export-onnx-ctc.py to get the onnx model.
+
+2. Run this file
+
+./zipformer/onnx_pretrained_ctc.py \
+  --nn-model /path/to/model.onnx \
+  --tokens /path/to/data/lang_bpe_500/tokens.txt \
+  1089-134686-0001.wav \
+  1221-135766-0001.wav \
+  1221-135766-0002.wav
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+import onnxruntime as ort
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="Path to the onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        nn_model: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_model(nn_model)
+
+    def init_model(self, nn_model: str):
+        self.model = ort.InferenceSession(
+            nn_model,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        meta = self.model.get_modelmeta().custom_metadata_map
+        print(meta)
+
+    def __call__(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D float tensor of shape (N, T, C)
+          x_lens:
+            A 1-D int64 tensor of shape (N,)
+        Returns:
+          Return a tuple containing:
+            - A float tensor containing log_probs of shape (N, T, C)
+            - A int64 tensor containing log_probs_len of shape (N)
+        """
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+                self.model.get_outputs()[1].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+                self.model.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        nn_model=args.nn_model,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    log_probs, log_probs_len = model(features, feature_lengths)
+
+    token_table = k2.SymbolTable.from_file(args.tokens)
+
+    def token_ids_to_words(token_ids: List[int]) -> str:
+        text = ""
+        for i in token_ids:
+            text += token_table[i]
+        return text.replace("▁", " ").strip()
+
+    blank_id = 0
+    s = "\n"
+    for i in range(log_probs.size(0)):
+        # greedy search
+        indexes = log_probs[i, : log_probs_len[i]].argmax(dim=-1)
+        token_ids = torch.unique_consecutive(indexes)
+
+        token_ids = token_ids[token_ids != blank_id]
+        words = token_ids_to_words(token_ids.tolist())
+        s += f"{args.sound_files[i]}:\n{words}\n\n"
+
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_H.py b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_H.py
new file mode 100755
index 000000000..683a7dc20
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_H.py
@@ -0,0 +1,277 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+"""
+This script loads ONNX models and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+as an example to show how to use this file.
+
+1. Please follow ./export-onnx-ctc.py to get the onnx model.
+
+2. Run this file
+
+./zipformer/onnx_pretrained_ctc_H.py \
+  --nn-model /path/to/model.onnx \
+  --tokens /path/to/data/lang_bpe_500/tokens.txt \
+  --H /path/to/H.fst \
+  1089-134686-0001.wav \
+  1221-135766-0001.wav \
+  1221-135766-0002.wav
+
+You can find exported ONNX models at
+https://huggingface.co/csukuangfj/sherpa-onnx-zipformer-ctc-en-2023-10-02
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+from typing import Dict
+import kaldifst
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="Path to the onnx model. ",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        help="""Path to tokens.txt.""",
+    )
+
+    parser.add_argument(
+        "--H",
+        type=str,
+        help="""Path to H.fst.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        nn_model: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_model(nn_model)
+
+    def init_model(self, nn_model: str):
+        self.model = ort.InferenceSession(
+            nn_model,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        meta = self.model.get_modelmeta().custom_metadata_map
+        print(meta)
+
+    def __call__(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D float tensor of shape (N, T, C)
+          x_lens:
+            A 1-D int64 tensor of shape (N,)
+        Returns:
+          Return a tuple containing:
+            - A float tensor containing log_probs of shape (N, T, C)
+            - A int64 tensor containing log_probs_len of shape (N)
+        """
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+                self.model.get_outputs()[1].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+                self.model.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    log_probs: torch.Tensor,
+    H: kaldifst,
+    id2token: Dict[int, str],
+) -> List[str]:
+    """
+    Args:
+      filename:
+        Path to the filename for decoding. Used for debugging.
+      log_probs:
+        A 2-D float32 tensor of shape (num_frames, vocab_size). It
+        contains output from log_softmax.
+      H:
+        The H graph.
+      id2word:
+        A map mapping token ID to word string.
+    Returns:
+      Return a list of decoded words.
+    """
+    logging.info(f"{filename}, {log_probs.shape}")
+    decodable = DecodableCtc(log_probs.cpu())
+
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(H, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        logging.info(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        logging.info(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # tokens are incremented during graph construction
+    # are shifted by 1 during graph construction
+    hyps = [id2token[i - 1] for i in osymbols_out if i != 1]
+    hyps = "".join(hyps).split("\u2581")  # unicode codepoint of ▁
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        nn_model=args.nn_model,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    logging.info(f"Loading H from {args.H}")
+    H = kaldifst.StdVectorFst.read(args.H)
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    log_probs, log_probs_len = model(features, feature_lengths)
+
+    token_table = k2.SymbolTable.from_file(args.tokens)
+
+    hyps = []
+    for i in range(log_probs.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[i],
+            log_probs=log_probs[i, : log_probs_len[i]],
+            H=H,
+            id2token=token_table,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HL.py b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HL.py
new file mode 100755
index 000000000..0b94bfa65
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HL.py
@@ -0,0 +1,275 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+"""
+This script loads ONNX models and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+as an example to show how to use this file.
+
+1. Please follow ./export-onnx-ctc.py to get the onnx model.
+
+2. Run this file
+
+./zipformer/onnx_pretrained_ctc_HL.py \
+  --nn-model /path/to/model.onnx \
+  --words /path/to/data/lang_bpe_500/words.txt \
+  --HL /path/to/HL.fst \
+  1089-134686-0001.wav \
+  1221-135766-0001.wav \
+  1221-135766-0002.wav
+
+You can find exported ONNX models at
+https://huggingface.co/csukuangfj/sherpa-onnx-zipformer-ctc-en-2023-10-02
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+from typing import Dict
+import kaldifst
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="Path to the onnx model. ",
+    )
+
+    parser.add_argument(
+        "--words",
+        type=str,
+        help="""Path to words.txt.""",
+    )
+
+    parser.add_argument(
+        "--HL",
+        type=str,
+        help="""Path to HL.fst.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        nn_model: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_model(nn_model)
+
+    def init_model(self, nn_model: str):
+        self.model = ort.InferenceSession(
+            nn_model,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        meta = self.model.get_modelmeta().custom_metadata_map
+        print(meta)
+
+    def __call__(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D float tensor of shape (N, T, C)
+          x_lens:
+            A 1-D int64 tensor of shape (N,)
+        Returns:
+          Return a tuple containing:
+            - A float tensor containing log_probs of shape (N, T, C)
+            - A int64 tensor containing log_probs_len of shape (N)
+        """
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+                self.model.get_outputs()[1].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+                self.model.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    log_probs: torch.Tensor,
+    HL: kaldifst,
+    id2word: Dict[int, str],
+) -> List[str]:
+    """
+    Args:
+      filename:
+        Path to the filename for decoding. Used for debugging.
+      log_probs:
+        A 2-D float32 tensor of shape (num_frames, vocab_size). It
+        contains output from log_softmax.
+      HL:
+        The HL graph.
+      id2word:
+        A map mapping word ID to word string.
+    Returns:
+      Return a list of decoded words.
+    """
+    logging.info(f"{filename}, {log_probs.shape}")
+    decodable = DecodableCtc(log_probs.cpu())
+
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(HL, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        logging.info(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        logging.info(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # are shifted by 1 during graph construction
+    hyps = [id2word[i] for i in osymbols_out]
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        nn_model=args.nn_model,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    logging.info(f"Loading HL from {args.HL}")
+    HL = kaldifst.StdVectorFst.read(args.HL)
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    log_probs, log_probs_len = model(features, feature_lengths)
+
+    word_table = k2.SymbolTable.from_file(args.words)
+
+    hyps = []
+    for i in range(log_probs.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[i],
+            log_probs=log_probs[i, : log_probs_len[i]],
+            HL=HL,
+            id2word=word_table,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py
new file mode 100755
index 000000000..93569142a
--- /dev/null
+++ b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py
@@ -0,0 +1,275 @@
+#!/usr/bin/env python3
+# Copyright      2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+"""
+This script loads ONNX models and uses them to decode waves.
+
+We use the pre-trained model from
+https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-transducer-ctc-2023-06-13
+as an example to show how to use this file.
+
+1. Please follow ./export-onnx-ctc.py to get the onnx model.
+
+2. Run this file
+
+./zipformer/onnx_pretrained_ctc_HLG.py \
+  --nn-model /path/to/model.onnx \
+  --words /path/to/data/lang_bpe_500/words.txt \
+  --HLG /path/to/HLG.fst \
+  1089-134686-0001.wav \
+  1221-135766-0001.wav \
+  1221-135766-0002.wav
+
+You can find exported ONNX models at
+https://huggingface.co/csukuangfj/sherpa-onnx-zipformer-ctc-en-2023-10-02
+"""
+
+import argparse
+import logging
+import math
+from typing import List, Tuple
+
+import k2
+import kaldifeat
+from typing import Dict
+import kaldifst
+import onnxruntime as ort
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="Path to the onnx model. ",
+    )
+
+    parser.add_argument(
+        "--words",
+        type=str,
+        help="""Path to words.txt.""",
+    )
+
+    parser.add_argument(
+        "--HLG",
+        type=str,
+        help="""Path to HLG.fst.""",
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    return parser
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        nn_model: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.init_model(nn_model)
+
+    def init_model(self, nn_model: str):
+        self.model = ort.InferenceSession(
+            nn_model,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        meta = self.model.get_modelmeta().custom_metadata_map
+        print(meta)
+
+    def __call__(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D float tensor of shape (N, T, C)
+          x_lens:
+            A 1-D int64 tensor of shape (N,)
+        Returns:
+          Return a tuple containing:
+            - A float tensor containing log_probs of shape (N, T, C)
+            - A int64 tensor containing log_probs_len of shape (N)
+        """
+        out = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+                self.model.get_outputs()[1].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+                self.model.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )
+        return torch.from_numpy(out[0]), torch.from_numpy(out[1])
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert (
+            sample_rate == expected_sample_rate
+        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    log_probs: torch.Tensor,
+    HLG: kaldifst,
+    id2word: Dict[int, str],
+) -> List[str]:
+    """
+    Args:
+      filename:
+        Path to the filename for decoding. Used for debugging.
+      log_probs:
+        A 2-D float32 tensor of shape (num_frames, vocab_size). It
+        contains output from log_softmax.
+      HLG:
+        The HLG graph.
+      id2word:
+        A map mapping word ID to word string.
+    Returns:
+      Return a list of decoded words.
+    """
+    logging.info(f"{filename}, {log_probs.shape}")
+    decodable = DecodableCtc(log_probs.cpu())
+
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(HLG, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        logging.info(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        logging.info(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # are shifted by 1 during graph construction
+    hyps = [id2word[i] for i in osymbols_out]
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+    model = OnnxModel(
+        nn_model=args.nn_model,
+    )
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = args.sample_rate
+    opts.mel_opts.num_bins = 80
+
+    logging.info(f"Loading HLG from {args.HLG}")
+    HLG = kaldifst.StdVectorFst.read(args.HLG)
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files,
+        expected_sample_rate=args.sample_rate,
+    )
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+    features = pad_sequence(
+        features,
+        batch_first=True,
+        padding_value=math.log(1e-10),
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
+    log_probs, log_probs_len = model(features, feature_lengths)
+
+    word_table = k2.SymbolTable.from_file(args.words)
+
+    hyps = []
+    for i in range(log_probs.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[i],
+            log_probs=log_probs[i, : log_probs_len[i]],
+            HLG=HLG,
+            id2word=word_table,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/librispeech/ASR/zipformer/optim.py b/egs/librispeech/ASR/zipformer/optim.py
index abfb2092c..8ee2b0eb4 100644
--- a/egs/librispeech/ASR/zipformer/optim.py
+++ b/egs/librispeech/ASR/zipformer/optim.py
@@ -116,7 +116,7 @@ class BatchedOptimizer(Optimizer):
 
         yield tuples  # <-- calling code will do the actual optimization here!
 
-        for ((stacked_params, _state, _names), batch) in zip(tuples, batches):
+        for (stacked_params, _state, _names), batch in zip(tuples, batches):
             for i, p in enumerate(batch):  # batch is list of Parameter
                 p.copy_(stacked_params[i])
 
@@ -181,7 +181,6 @@ class ScaledAdam(BatchedOptimizer):
         size_update_period=4,
         clipping_update_period=100,
     ):
-
         defaults = dict(
             lr=lr,
             clipping_scale=clipping_scale,
@@ -299,8 +298,8 @@ class ScaledAdam(BatchedOptimizer):
             # the input is groups of parameter or named parameter.
             for cur_group in iterable_or_groups:
                 assert "named_params" in cur_group
-                name_list = [ x[0] for x in cur_group["named_params"] ]
-                p_list = [ x[1] for x in cur_group["named_params"] ]
+                name_list = [x[0] for x in cur_group["named_params"]]
+                p_list = [x[1] for x in cur_group["named_params"]]
                 del cur_group["named_params"]
                 cur_group["params"] = p_list
                 param_groups.append(cur_group)
@@ -327,9 +326,7 @@ class ScaledAdam(BatchedOptimizer):
         batch = True
 
         for group, group_params_names in zip(self.param_groups, self.parameters_names):
-
             with self.batched_params(group["params"], group_params_names) as batches:
-
                 # batches is list of pairs (stacked_param, state).  stacked_param is like
                 # a regular parameter, and will have a .grad, but the 1st dim corresponds to
                 # a stacking dim, it is not a real dim.
@@ -428,7 +425,7 @@ class ScaledAdam(BatchedOptimizer):
         clipping_update_period = group["clipping_update_period"]
 
         tot_sumsq = torch.tensor(0.0, device=first_p.device)
-        for (p, state, param_names) in tuples:
+        for p, state, param_names in tuples:
             grad = p.grad
             if grad.is_sparse:
                 raise RuntimeError(
@@ -494,6 +491,12 @@ class ScaledAdam(BatchedOptimizer):
                 if self.show_dominant_parameters:
                     assert p.shape[0] == len(param_names)
                     self._show_gradient_dominating_parameter(tuples, tot_sumsq)
+            if ans != ans:  # e.g. ans is nan
+                ans = 0.0
+            if ans == 0.0:
+                for p, state, param_names in tuples:
+                    p.grad.zero_()  # get rid of infinity()
+
             return ans
 
     def _show_gradient_dominating_parameter(
@@ -513,7 +516,7 @@ class ScaledAdam(BatchedOptimizer):
                 from tuples, we still pass it to save some time.
         """
         all_sumsq_orig = {}
-        for (p, state, batch_param_names) in tuples:
+        for p, state, batch_param_names in tuples:
             # p is a stacked batch parameters.
             batch_grad = p.grad
             if p.numel() == p.shape[0]:  # a batch of scalars
@@ -529,7 +532,6 @@ class ScaledAdam(BatchedOptimizer):
             for name, sumsq_orig, rms, grad in zip(
                 batch_param_names, batch_sumsq_orig, batch_rms_orig, batch_grad
             ):
-
                 proportion_orig = sumsq_orig / tot_sumsq
                 all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad)
 
@@ -577,7 +579,7 @@ class ScaledAdam(BatchedOptimizer):
 
         grad = p.grad
         if clipping_scale != 1.0:
-            grad = grad * clipping_scale
+            grad *= clipping_scale
         step = state["step"]
         delta = state["delta"]
 
@@ -667,8 +669,7 @@ class ScaledAdam(BatchedOptimizer):
         # We have to look at the trained model for parameters at or around the
         # param_max_rms, because sometimes they can indicate a problem with the
         # topology or settings.
-        scale_step = torch.minimum(scale_step,
-                                   (param_max_rms - param_rms) / param_rms)
+        scale_step = torch.minimum(scale_step, (param_max_rms - param_rms) / param_rms)
 
         delta = state["delta"]
         # the factor of (1-beta1) relates to momentum.
@@ -879,7 +880,8 @@ class Eden(LRScheduler):
         warmup_factor = (
             1.0
             if self.batch >= self.warmup_batches
-            else self.warmup_start + (1.0 - self.warmup_start) * (self.batch / self.warmup_batches)
+            else self.warmup_start
+            + (1.0 - self.warmup_start) * (self.batch / self.warmup_batches)
             # else 0.5 + 0.5 * (self.batch / self.warmup_batches)
         )
 
@@ -1111,7 +1113,7 @@ def _test_scaled_adam(hidden_dim: int):
 
             # if epoch == 130:
             #    opts = diagnostics.TensorDiagnosticOptions(
-            #        2 ** 22
+            #        512
             #    )  # allow 4 megabytes per sub-module
             #    diagnostic = diagnostics.attach_diagnostics(m, opts)
 
diff --git a/egs/librispeech/ASR/zipformer/profile.py b/egs/librispeech/ASR/zipformer/profile.py
index b460b5338..57f44a90a 100755
--- a/egs/librispeech/ASR/zipformer/profile.py
+++ b/egs/librispeech/ASR/zipformer/profile.py
@@ -100,17 +100,13 @@ class Model(nn.Module):
         self.encoder_embed = encoder_embed
         self.encoder_proj = encoder_proj
 
-    def forward(
-        self, feature: Tensor, feature_lens: Tensor
-    ) -> Tuple[Tensor, Tensor]:
+    def forward(self, feature: Tensor, feature_lens: Tensor) -> Tuple[Tensor, Tensor]:
         x, x_lens = self.encoder_embed(feature, feature_lens)
 
         src_key_padding_mask = make_pad_mask(x_lens)
         x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
 
-        encoder_out, encoder_out_lens = self.encoder(
-            x, x_lens, src_key_padding_mask
-        )
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask)
 
         encoder_out = encoder_out.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
         logits = self.encoder_proj(encoder_out)
@@ -168,9 +164,7 @@ def main():
 
 
 if __name__ == "__main__":
-    formatter = (
-        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
-    )
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
     logging.basicConfig(format=formatter, level=logging.INFO)
 
     main()
diff --git a/egs/librispeech/ASR/zipformer/scaling.py b/egs/librispeech/ASR/zipformer/scaling.py
index 7c98ef045..c0f1e3087 100644
--- a/egs/librispeech/ASR/zipformer/scaling.py
+++ b/egs/librispeech/ASR/zipformer/scaling.py
@@ -25,6 +25,7 @@ import math
 import torch.nn as nn
 from torch import Tensor
 
+
 def logaddexp_onnx(x: Tensor, y: Tensor) -> Tensor:
     max_value = torch.max(x, y)
     diff = torch.abs(x - y)
@@ -55,28 +56,34 @@ def logaddexp(x: Tensor, y: Tensor) -> Tensor:
         # for torch.jit.trace()
         return torch.logaddexp(x, y)
 
+
 class PiecewiseLinear(object):
     """
     Piecewise linear function, from float to float, specified as nonempty list of (x,y) pairs with
     the x values in order.  x values <[initial x] or >[final x] are map to [initial y], [final y]
     respectively.
     """
+
     def __init__(self, *args):
         assert len(args) >= 1, len(args)
         if len(args) == 1 and isinstance(args[0], PiecewiseLinear):
             self.pairs = list(args[0].pairs)
         else:
-            self.pairs = [ (float(x), float(y)) for x,y in args ]
-        for (x,y) in self.pairs:
+            self.pairs = [(float(x), float(y)) for x, y in args]
+        for x, y in self.pairs:
             assert isinstance(x, (float, int)), type(x)
             assert isinstance(y, (float, int)), type(y)
 
         for i in range(len(self.pairs) - 1):
-            assert self.pairs[i + 1][0] > self.pairs[i][0], (i, self.pairs[i], self.pairs[i + 1])
+            assert self.pairs[i + 1][0] > self.pairs[i][0], (
+                i,
+                self.pairs[i],
+                self.pairs[i + 1],
+            )
 
     def __str__(self):
         # e.g. 'PiecewiseLinear((0., 10.), (100., 0.))'
-        return f'PiecewiseLinear({str(self.pairs)[1:-1]})'
+        return f"PiecewiseLinear({str(self.pairs)[1:-1]})"
 
     def __call__(self, x):
         if x <= self.pairs[0][0]:
@@ -93,37 +100,36 @@ class PiecewiseLinear(object):
             assert False
 
     def __mul__(self, alpha):
-        return PiecewiseLinear(
-            * [(x, y * alpha) for x, y in self.pairs])
+        return PiecewiseLinear(*[(x, y * alpha) for x, y in self.pairs])
 
     def __add__(self, x):
         if isinstance(x, (float, int)):
-            return PiecewiseLinear(
-                * [(p[0], p[1] + x) for p in self.pairs])
+            return PiecewiseLinear(*[(p[0], p[1] + x) for p in self.pairs])
         s, x = self.get_common_basis(x)
         return PiecewiseLinear(
-            * [(sp[0], sp[1] + xp[1]) for sp, xp in zip(s.pairs, x.pairs)])
+            *[(sp[0], sp[1] + xp[1]) for sp, xp in zip(s.pairs, x.pairs)]
+        )
 
     def max(self, x):
         if isinstance(x, (float, int)):
-            x = PiecewiseLinear( (0, x) )
+            x = PiecewiseLinear((0, x))
         s, x = self.get_common_basis(x, include_crossings=True)
         return PiecewiseLinear(
-            * [(sp[0], max(sp[1], xp[1])) for sp, xp in zip(s.pairs, x.pairs)])
+            *[(sp[0], max(sp[1], xp[1])) for sp, xp in zip(s.pairs, x.pairs)]
+        )
 
     def min(self, x):
         if isinstance(x, float) or isinstance(x, int):
-            x = PiecewiseLinear( (0, x) )
+            x = PiecewiseLinear((0, x))
         s, x = self.get_common_basis(x, include_crossings=True)
         return PiecewiseLinear(
-            * [ (sp[0], min(sp[1], xp[1])) for sp, xp in zip(s.pairs, x.pairs)])
+            *[(sp[0], min(sp[1], xp[1])) for sp, xp in zip(s.pairs, x.pairs)]
+        )
 
     def __eq__(self, other):
         return self.pairs == other.pairs
 
-    def get_common_basis(self,
-                         p: 'PiecewiseLinear',
-                         include_crossings: bool = False):
+    def get_common_basis(self, p: "PiecewiseLinear", include_crossings: bool = False):
         """
         Returns (self_mod, p_mod) which are equivalent piecewise linear
         functions to self and p, but with the same x values.
@@ -135,28 +141,30 @@ class PiecewiseLinear(object):
         assert isinstance(p, PiecewiseLinear), type(p)
 
         # get sorted x-values without repetition.
-        x_vals = sorted(set([ x for x, _ in self.pairs ] + [ x for x, _ in p.pairs ]))
-        y_vals1 = [ self(x) for x in x_vals ]
-        y_vals2 = [ p(x) for x in x_vals ]
+        x_vals = sorted(set([x for x, _ in self.pairs] + [x for x, _ in p.pairs]))
+        y_vals1 = [self(x) for x in x_vals]
+        y_vals2 = [p(x) for x in x_vals]
 
         if include_crossings:
             extra_x_vals = []
             for i in range(len(x_vals) - 1):
-                if (y_vals1[i] > y_vals2[i]) != (y_vals1[i+1] > y_vals2[i+1]):
+                if (y_vals1[i] > y_vals2[i]) != (y_vals1[i + 1] > y_vals2[i + 1]):
                     # if the two lines in this subsegment potentially cross each other..
                     diff_cur = abs(y_vals1[i] - y_vals2[i])
-                    diff_next = abs(y_vals1[i+1] - y_vals2[i+1])
+                    diff_next = abs(y_vals1[i + 1] - y_vals2[i + 1])
                     # `pos`, between 0 and 1, gives the relative x position,
                     # with 0 being x_vals[i] and 1 being x_vals[i+1].
                     pos = diff_cur / (diff_cur + diff_next)
-                    extra_x_val = x_vals[i] + pos * (x_vals[i+1] - x_vals[i])
+                    extra_x_val = x_vals[i] + pos * (x_vals[i + 1] - x_vals[i])
                     extra_x_vals.append(extra_x_val)
             if len(extra_x_vals) > 0:
                 x_vals = sorted(set(x_vals + extra_x_vals))
-        y_vals1 = [ self(x) for x in x_vals ]
-        y_vals2 = [ p(x) for x in x_vals ]
-        return ( PiecewiseLinear(* zip(x_vals, y_vals1)),
-                 PiecewiseLinear(* zip(x_vals, y_vals2)) )
+        y_vals1 = [self(x) for x in x_vals]
+        y_vals2 = [p(x) for x in x_vals]
+        return (
+            PiecewiseLinear(*zip(x_vals, y_vals1)),
+            PiecewiseLinear(*zip(x_vals, y_vals2)),
+        )
 
 
 class ScheduledFloat(torch.nn.Module):
@@ -176,9 +184,8 @@ class ScheduledFloat(torch.nn.Module):
     `default` is used when self.batch_count is not set or not in training mode or in
      torch.jit scripting mode.
     """
-    def __init__(self,
-                 *args,
-                 default: float = 0.0):
+
+    def __init__(self, *args, default: float = 0.0):
         super().__init__()
         # self.batch_count and self.name will be written to in the training loop.
         self.batch_count = None
@@ -187,47 +194,55 @@ class ScheduledFloat(torch.nn.Module):
         self.schedule = PiecewiseLinear(*args)
 
     def extra_repr(self) -> str:
-        return f'batch_count={self.batch_count}, schedule={str(self.schedule.pairs[1:-1])}'
+        return (
+            f"batch_count={self.batch_count}, schedule={str(self.schedule.pairs[1:-1])}"
+        )
 
     def __float__(self):
         batch_count = self.batch_count
-        if batch_count is None or not self.training or torch.jit.is_scripting() or torch.jit.is_tracing():
+        if (
+            batch_count is None
+            or not self.training
+            or torch.jit.is_scripting()
+            or torch.jit.is_tracing()
+        ):
             return float(self.default)
         else:
             ans = self.schedule(self.batch_count)
             if random.random() < 0.0002:
-                logging.info(f"ScheduledFloat: name={self.name}, batch_count={self.batch_count}, ans={ans}")
+                logging.info(
+                    f"ScheduledFloat: name={self.name}, batch_count={self.batch_count}, ans={ans}"
+                )
             return ans
 
     def __add__(self, x):
         if isinstance(x, float) or isinstance(x, int):
-            return ScheduledFloat(self.schedule + x,
-                                  default=self.default)
+            return ScheduledFloat(self.schedule + x, default=self.default)
         else:
-            return ScheduledFloat(self.schedule + x.schedule,
-                                  default=self.default+x.default)
+            return ScheduledFloat(
+                self.schedule + x.schedule, default=self.default + x.default
+            )
 
     def max(self, x):
         if isinstance(x, float) or isinstance(x, int):
-            return ScheduledFloat(self.schedule.max(x),
-                                  default=self.default)
+            return ScheduledFloat(self.schedule.max(x), default=self.default)
         else:
-            return ScheduledFloat(self.schedule.max(x.schedule),
-                                  default=max(self.default, x.default))
+            return ScheduledFloat(
+                self.schedule.max(x.schedule), default=max(self.default, x.default)
+            )
 
 
 FloatLike = Union[float, ScheduledFloat]
 
 
-def random_cast_to_half(x: Tensor,
-                        min_abs: float = 5.0e-06) -> Tensor:
+def random_cast_to_half(x: Tensor, min_abs: float = 5.0e-06) -> Tensor:
     """
     A randomized way of casting a floating point value to half precision.
     """
     if x.dtype == torch.float16:
         return x
     x_abs = x.abs()
-    is_too_small = (x_abs < min_abs)
+    is_too_small = x_abs < min_abs
     # for elements where is_too_small is true, random_val will contain +-min_abs with
     # probability (x.abs() / min_abs), and 0.0 otherwise.  [so this preserves expectations,
     # for those elements].
@@ -242,6 +257,7 @@ class CutoffEstimator:
 
       p is the proportion of items that should be above the cutoff.
     """
+
     def __init__(self, p: float):
         self.p = p
         # total count of items
@@ -255,7 +271,7 @@ class CutoffEstimator:
         """
         Returns true if x is above the cutoff.
         """
-        ans = (x > self.cutoff)
+        ans = x > self.cutoff
         self.count += 1
         if ans:
             self.count_above += 1
@@ -263,7 +279,7 @@ class CutoffEstimator:
         delta_p = cur_p - self.p
         if (delta_p > 0) == ans:
             q = abs(delta_p)
-            self.cutoff = x * q + self.cutoff * (1-q)
+            self.cutoff = x * q + self.cutoff * (1 - q)
         return ans
 
 
@@ -272,6 +288,7 @@ class SoftmaxFunction(torch.autograd.Function):
     Tries to handle half-precision derivatives in a randomized way that should
     be more accurate for training than the default behavior.
     """
+
     @staticmethod
     def forward(ctx, x: Tensor, dim: int):
         ans = x.softmax(dim=dim)
@@ -287,7 +304,7 @@ class SoftmaxFunction(torch.autograd.Function):
 
     @staticmethod
     def backward(ctx, ans_grad: Tensor):
-        ans, = ctx.saved_tensors
+        (ans,) = ctx.saved_tensors
         with torch.cuda.amp.autocast(enabled=False):
             ans_grad = ans_grad.to(torch.float32)
             ans = ans.to(torch.float32)
@@ -306,17 +323,16 @@ def softmax(x: Tensor, dim: int):
 class MaxEigLimiterFunction(torch.autograd.Function):
     @staticmethod
     def forward(
-            ctx,
-            x: Tensor,
-            coeffs: Tensor,
-            direction: Tensor,
-            channel_dim: int,
-            grad_scale: float) -> Tensor:
+        ctx,
+        x: Tensor,
+        coeffs: Tensor,
+        direction: Tensor,
+        channel_dim: int,
+        grad_scale: float,
+    ) -> Tensor:
         ctx.channel_dim = channel_dim
         ctx.grad_scale = grad_scale
-        ctx.save_for_backward(x.detach(),
-                              coeffs.detach(),
-                              direction.detach())
+        ctx.save_for_backward(x.detach(), coeffs.detach(), direction.detach())
         return x
 
     @staticmethod
@@ -328,15 +344,20 @@ class MaxEigLimiterFunction(torch.autograd.Function):
             x = x_orig.transpose(ctx.channel_dim, -1).reshape(-1, num_channels)
             new_direction.requires_grad = False
             x = x - x.mean(dim=0)
-            x_var = (x ** 2).mean()
+            x_var = (x**2).mean()
             x_residual = x - coeffs * new_direction
-            x_residual_var = (x_residual ** 2).mean()
+            x_residual_var = (x_residual**2).mean()
             # `variance_proportion` is the proportion of the variance accounted for
             # by the top eigen-direction.  This is to be minimized.
             variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20)
             variance_proportion.backward()
         x_orig_grad = x_orig.grad
-        x_extra_grad = x_orig.grad * ctx.grad_scale * x_grad.norm() / (x_orig_grad.norm() + 1.0e-20)
+        x_extra_grad = (
+            x_orig.grad
+            * ctx.grad_scale
+            * x_grad.norm()
+            / (x_orig_grad.norm() + 1.0e-20)
+        )
         return x_grad + x_extra_grad.detach(), None, None, None, None
 
 
@@ -348,8 +369,14 @@ class BiasNormFunction(torch.autograd.Function):
     # it can just store the returned value (chances are, this will also be needed for
     # some other reason, related to the next operation, so we can save memory).
     @staticmethod
-    def forward(ctx, x: Tensor, bias: Tensor, log_scale: Tensor, channel_dim: int,
-                store_output_for_backprop: bool) -> Tensor:
+    def forward(
+        ctx,
+        x: Tensor,
+        bias: Tensor,
+        log_scale: Tensor,
+        channel_dim: int,
+        store_output_for_backprop: bool,
+    ) -> Tensor:
         assert bias.ndim == 1
         if channel_dim < 0:
             channel_dim = channel_dim + x.ndim
@@ -357,10 +384,16 @@ class BiasNormFunction(torch.autograd.Function):
         ctx.channel_dim = channel_dim
         for _ in range(channel_dim + 1, x.ndim):
             bias = bias.unsqueeze(-1)
-        scales = (torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) ** -0.5) * log_scale.exp()
+        scales = (
+            torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) ** -0.5
+        ) * log_scale.exp()
         ans = x * scales
-        ctx.save_for_backward(ans.detach() if store_output_for_backprop else x,
-                              scales.detach(), bias.detach(), log_scale.detach())
+        ctx.save_for_backward(
+            ans.detach() if store_output_for_backprop else x,
+            scales.detach(),
+            bias.detach(),
+            log_scale.detach(),
+        )
         return ans
 
     @staticmethod
@@ -376,7 +409,9 @@ class BiasNormFunction(torch.autograd.Function):
         log_scale.requires_grad = True
         with torch.enable_grad():
             # recompute scales from x, bias and log_scale.
-            scales = (torch.mean((x - bias) ** 2, dim=ctx.channel_dim, keepdim=True) ** -0.5) * log_scale.exp()
+            scales = (
+                torch.mean((x - bias) ** 2, dim=ctx.channel_dim, keepdim=True) ** -0.5
+            ) * log_scale.exp()
             ans = x * scales
             ans.backward(gradient=ans_grad)
         return x.grad, bias.grad.flatten(), log_scale.grad, None, None
@@ -412,14 +447,15 @@ class BiasNorm(torch.nn.Module):
          than the input of this module to be required to be stored for the
          backprop.
     """
+
     def __init__(
-            self,
-            num_channels: int,
-            channel_dim: int = -1,  # CAUTION: see documentation.
-            log_scale: float = 1.0,
-            log_scale_min: float = -1.5,
-            log_scale_max: float = 1.5,
-            store_output_for_backprop: bool = False
+        self,
+        num_channels: int,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+        log_scale: float = 1.0,
+        log_scale_min: float = -1.5,
+        log_scale_max: float = 1.5,
+        store_output_for_backprop: bool = False,
     ) -> None:
         super(BiasNorm, self).__init__()
         self.num_channels = num_channels
@@ -442,23 +478,24 @@ class BiasNorm(torch.nn.Module):
             bias = self.bias
             for _ in range(channel_dim + 1, x.ndim):
                 bias = bias.unsqueeze(-1)
-            scales = ((torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) ** -0.5) *
-                      self.log_scale.exp())
+            scales = (
+                torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) ** -0.5
+            ) * self.log_scale.exp()
             return x * scales
 
-        log_scale = limit_param_value(self.log_scale,
-                                      min=float(self.log_scale_min),
-                                      max=float(self.log_scale_max),
-                                      training=self.training)
+        log_scale = limit_param_value(
+            self.log_scale,
+            min=float(self.log_scale_min),
+            max=float(self.log_scale_max),
+            training=self.training,
+        )
 
-        return BiasNormFunction.apply(x, self.bias, log_scale,
-                                      self.channel_dim,
-                                      self.store_output_for_backprop)
+        return BiasNormFunction.apply(
+            x, self.bias, log_scale, self.channel_dim, self.store_output_for_backprop
+        )
 
 
-def ScaledLinear(*args,
-                 initial_scale: float = 1.0,
-                 **kwargs) -> nn.Linear:
+def ScaledLinear(*args, initial_scale: float = 1.0, **kwargs) -> nn.Linear:
     """
     Behaves like a constructor of a modified version of nn.Linear
     that gives an easy way to set the default initial parameter scale.
@@ -477,15 +514,11 @@ def ScaledLinear(*args,
     with torch.no_grad():
         ans.weight[:] *= initial_scale
         if ans.bias is not None:
-            torch.nn.init.uniform_(ans.bias,
-                                   -0.1 * initial_scale,
-                                   0.1 * initial_scale)
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
     return ans
 
 
-def ScaledConv1d(*args,
-                 initial_scale: float = 1.0,
-                 **kwargs) -> nn.Conv1d:
+def ScaledConv1d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv1d:
     """
     Behaves like a constructor of a modified version of nn.Conv1d
     that gives an easy way to set the default initial parameter scale.
@@ -504,15 +537,11 @@ def ScaledConv1d(*args,
     with torch.no_grad():
         ans.weight[:] *= initial_scale
         if ans.bias is not None:
-            torch.nn.init.uniform_(ans.bias,
-                                   -0.1 * initial_scale,
-                                   0.1 * initial_scale)
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
     return ans
 
 
-def ScaledConv2d(*args,
-                 initial_scale: float = 1.0,
-                 **kwargs) -> nn.Conv2d:
+def ScaledConv2d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv2d:
     """
     Behaves like a constructor of a modified version of nn.Conv2d
     that gives an easy way to set the default initial parameter scale.
@@ -532,9 +561,7 @@ def ScaledConv2d(*args,
     with torch.no_grad():
         ans.weight[:] *= initial_scale
         if ans.bias is not None:
-            torch.nn.init.uniform_(ans.bias,
-                                   -0.1 * initial_scale,
-                                   0.1 * initial_scale)
+            torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale)
     return ans
 
 
@@ -562,29 +589,36 @@ class ChunkCausalDepthwiseConv1d(torch.nn.Module):
            Another option, if you want to do something like this, is
            to re-initialize the parameters.
     """
-    def __init__(self,
-                 channels: int,
-                 kernel_size: int,
-                 initial_scale: float = 1.0,
-                 bias: bool = True):
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int,
+        initial_scale: float = 1.0,
+        bias: bool = True,
+    ):
         super().__init__()
         assert kernel_size % 2 == 1
 
         half_kernel_size = (kernel_size + 1) // 2
         # will pad manually, on one side.
-        self.causal_conv = nn.Conv1d(in_channels=channels,
-                                     out_channels=channels,
-                                     groups=channels,
-                                     kernel_size=half_kernel_size,
-                                     padding=0,
-                                     bias=True)
+        self.causal_conv = nn.Conv1d(
+            in_channels=channels,
+            out_channels=channels,
+            groups=channels,
+            kernel_size=half_kernel_size,
+            padding=0,
+            bias=True,
+        )
 
-        self.chunkwise_conv = nn.Conv1d(in_channels=channels,
-                                        out_channels=channels,
-                                        groups=channels,
-                                        kernel_size=kernel_size,
-                                        padding=kernel_size // 2,
-                                        bias=bias)
+        self.chunkwise_conv = nn.Conv1d(
+            in_channels=channels,
+            out_channels=channels,
+            groups=channels,
+            kernel_size=kernel_size,
+            padding=kernel_size // 2,
+            bias=bias,
+        )
 
         # first row is correction factors added to the scale near the left edge of the chunk,
         # second row is correction factors added to the scale near the right edge of the chunk,
@@ -596,17 +630,15 @@ class ChunkCausalDepthwiseConv1d(torch.nn.Module):
             self.causal_conv.weight[:] *= initial_scale
             self.chunkwise_conv.weight[:] *= initial_scale
             if bias:
-                torch.nn.init.uniform_(self.causal_conv.bias,
-                                       -0.1 * initial_scale,
-                                       0.1 * initial_scale)
+                torch.nn.init.uniform_(
+                    self.causal_conv.bias, -0.1 * initial_scale, 0.1 * initial_scale
+                )
 
-    def forward(self,
-                x: Tensor,
-                chunk_size: int = -1) -> Tensor:
+    def forward(self, x: Tensor, chunk_size: int = -1) -> Tensor:
         """
-        Forward function.  Args:
-          x: a Tensor of shape (batch_size, channels, seq_len)
-   chunk_size: the chunk size, in frames; does not have to divide seq_len exactly.
+             Forward function.  Args:
+               x: a Tensor of shape (batch_size, channels, seq_len)
+        chunk_size: the chunk size, in frames; does not have to divide seq_len exactly.
         """
         (batch_size, num_channels, seq_len) = x.shape
 
@@ -622,28 +654,32 @@ class ChunkCausalDepthwiseConv1d(torch.nn.Module):
 
         x = torch.nn.functional.pad(x, (left_pad, right_pad))
 
-        x_causal = self.causal_conv(x[..., :left_pad + seq_len])
+        x_causal = self.causal_conv(x[..., : left_pad + seq_len])
         assert x_causal.shape == (batch_size, num_channels, seq_len)
 
         x_chunk = x[..., left_pad:]
         num_chunks = x_chunk.shape[2] // chunk_size
         x_chunk = x_chunk.reshape(batch_size, num_channels, num_chunks, chunk_size)
-        x_chunk = x_chunk.permute(0, 2, 1, 3).reshape(batch_size * num_chunks,
-                                                      num_channels, chunk_size)
+        x_chunk = x_chunk.permute(0, 2, 1, 3).reshape(
+            batch_size * num_chunks, num_channels, chunk_size
+        )
         x_chunk = self.chunkwise_conv(x_chunk)  # does not change shape
 
         chunk_scale = self._get_chunk_scale(chunk_size)
 
         x_chunk = x_chunk * chunk_scale
-        x_chunk = x_chunk.reshape(batch_size, num_chunks,
-                                  num_channels, chunk_size).permute(0, 2, 1, 3)
-        x_chunk = x_chunk.reshape(batch_size, num_channels, num_chunks * chunk_size)[..., :seq_len]
+        x_chunk = x_chunk.reshape(
+            batch_size, num_chunks, num_channels, chunk_size
+        ).permute(0, 2, 1, 3)
+        x_chunk = x_chunk.reshape(batch_size, num_channels, num_chunks * chunk_size)[
+            ..., :seq_len
+        ]
 
         return x_chunk + x_causal
 
     def _get_chunk_scale(self, chunk_size: int):
         """Returns tensor of shape (num_channels, chunk_size) that will be used to
-           scale the output of self.chunkwise_conv."""
+        scale the output of self.chunkwise_conv."""
         left_edge = self.chunkwise_conv_scale[0]
         right_edge = self.chunkwise_conv_scale[1]
         if chunk_size < self.kernel_size:
@@ -652,9 +688,9 @@ class ChunkCausalDepthwiseConv1d(torch.nn.Module):
         else:
             t = chunk_size - self.kernel_size
             channels = left_edge.shape[0]
-            pad = torch.zeros(channels, t,
-                              device=left_edge.device,
-                              dtype=left_edge.dtype)
+            pad = torch.zeros(
+                channels, t, device=left_edge.device, dtype=left_edge.dtype
+            )
             left_edge = torch.cat((left_edge, pad), dim=-1)
             right_edge = torch.cat((pad, right_edge), dim=-1)
         return 1.0 + (left_edge + right_edge)
@@ -698,14 +734,14 @@ class ChunkCausalDepthwiseConv1d(torch.nn.Module):
 class BalancerFunction(torch.autograd.Function):
     @staticmethod
     def forward(
-            ctx,
-            x: Tensor,
-            min_mean: float,
-            max_mean: float,
-            min_rms: float,
-            max_rms: float,
-            grad_scale: float,
-            channel_dim: int,
+        ctx,
+        x: Tensor,
+        min_mean: float,
+        max_mean: float,
+        min_rms: float,
+        max_rms: float,
+        grad_scale: float,
+        channel_dim: int,
     ) -> Tensor:
         if channel_dim < 0:
             channel_dim += x.ndim
@@ -715,10 +751,8 @@ class BalancerFunction(torch.autograd.Function):
         return x
 
     @staticmethod
-    def backward(
-        ctx, x_grad: Tensor
-    ) -> Tuple[Tensor, None, None, None, None, None]:
-        x, = ctx.saved_tensors
+    def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None, None, None]:
+        (x,) = ctx.saved_tensors
         (min_mean, max_mean, min_rms, max_rms, grad_scale, channel_dim) = ctx.config
 
         try:
@@ -727,8 +761,8 @@ class BalancerFunction(torch.autograd.Function):
                     x = x.to(torch.float32)
                     x = x.detach()
                     x.requires_grad = True
-                    mean_dims = [ i for i in range(x.ndim) if i != channel_dim ]
-                    uncentered_var = (x ** 2).mean(dim=mean_dims, keepdim=True)
+                    mean_dims = [i for i in range(x.ndim) if i != channel_dim]
+                    uncentered_var = (x**2).mean(dim=mean_dims, keepdim=True)
                     mean = x.mean(dim=mean_dims, keepdim=True)
                     stddev = (uncentered_var - (mean * mean)).clamp(min=1.0e-20).sqrt()
                     rms = uncentered_var.clamp(min=1.0e-20).sqrt()
@@ -742,11 +776,16 @@ class BalancerFunction(torch.autograd.Function):
                     rms_clamped = rms.clamp(min=min_rms, max=max_rms)
                     r_loss = (rms_clamped / rms).log().abs()
 
-                    loss = (m_loss + r_loss)
+                    loss = m_loss + r_loss
 
                     loss.backward(gradient=torch.ones_like(loss))
                     loss_grad = x.grad
-                    loss_grad_rms = (loss_grad ** 2).mean(dim=mean_dims, keepdim=True).sqrt().clamp(min=1.0e-20)
+                    loss_grad_rms = (
+                        (loss_grad**2)
+                        .mean(dim=mean_dims, keepdim=True)
+                        .sqrt()
+                        .clamp(min=1.0e-20)
+                    )
 
                     loss_grad = loss_grad * (grad_scale / loss_grad_rms)
 
@@ -757,7 +796,9 @@ class BalancerFunction(torch.autograd.Function):
                     x_grad_mod = x_grad_float + (x_grad_float.abs() * loss_grad)
                     x_grad = x_grad_mod.to(x_grad.dtype)
         except Exception as e:
-            logging.info(f"Caught exception in Balancer backward: {e}, size={list(x_grad.shape)}, will continue.")
+            logging.info(
+                f"Caught exception in Balancer backward: {e}, size={list(x_grad.shape)}, will continue."
+            )
 
         return x_grad, None, None, None, None, None, None
 
@@ -793,16 +834,17 @@ class Balancer(torch.nn.Module):
              on each forward().  This is done randomly to prevent all layers
              from doing it at the same time.
     """
+
     def __init__(
-            self,
-            num_channels: int,
-            channel_dim: int,
-            min_positive: FloatLike = 0.05,
-            max_positive: FloatLike = 0.95,
-            min_abs: FloatLike = 0.2,
-            max_abs: FloatLike = 100.0,
-            grad_scale: FloatLike = 0.04,
-            prob: Optional[FloatLike] = None,
+        self,
+        num_channels: int,
+        channel_dim: int,
+        min_positive: FloatLike = 0.05,
+        max_positive: FloatLike = 0.95,
+        min_abs: FloatLike = 0.2,
+        max_abs: FloatLike = 100.0,
+        grad_scale: FloatLike = 0.04,
+        prob: Optional[FloatLike] = None,
     ):
         super().__init__()
 
@@ -823,8 +865,11 @@ class Balancer(torch.nn.Module):
         self.grad_scale = grad_scale
 
     def forward(self, x: Tensor) -> Tensor:
-        if (torch.jit.is_scripting() or not x.requires_grad or
-           (x.is_cuda and self.mem_cutoff(torch.cuda.memory_allocated()))):
+        if (
+            torch.jit.is_scripting()
+            or not x.requires_grad
+            or (x.is_cuda and self.mem_cutoff(torch.cuda.memory_allocated()))
+        ):
             return _no_op(x)
 
         prob = float(self.prob)
@@ -842,7 +887,7 @@ class Balancer(torch.nn.Module):
                     eps = 1.0e-10
                     # eps is to prevent crashes if x is exactly 0 or 1.
                     # we'll just end up returning a fairly large value.
-                    return (math.log (1+x+eps) - math.log (1-x+eps)) / 2.
+                    return (math.log(1 + x + eps) - math.log(1 - x + eps)) / 2.0
 
                 def _approx_inverse_erf(x):
                     # 1 / (sqrt(pi) * ln(2)),
@@ -853,6 +898,7 @@ class Balancer(torch.nn.Module):
                     # and math.erf(0.0407316414078772) = 0.045935330944660666,
                     # which is pretty close to 0.05.
                     return 0.8139535143 * _atanh(x)
+
                 # first convert x from the range 0..1 to the range -1..1 which the error
                 # function returns
                 x = -1 + (2 * x)
@@ -873,8 +919,9 @@ class Balancer(torch.nn.Module):
             return _no_op(x)
 
 
-def penalize_abs_values_gt(x: Tensor, limit: float, penalty: float,
-                           name: str = None) -> Tensor:
+def penalize_abs_values_gt(
+    x: Tensor, limit: float, penalty: float, name: str = None
+) -> Tensor:
     """
     Returns x unmodified, but in backprop will put a penalty for the excess of
     the absolute values of elements of x over the limit "limit".  E.g. if
@@ -910,13 +957,12 @@ def _diag(x: Tensor):  # like .diag(), but works for tensors with 3 dims.
     else:
         (batch, dim, dim) = x.shape
         x = x.reshape(batch, dim * dim)
-        x = x[:, ::dim+1]
+        x = x[:, :: dim + 1]
         assert x.shape == (batch, dim)
         return x
 
 
-def _whitening_metric(x: Tensor,
-                      num_groups: int):
+def _whitening_metric(x: Tensor, num_groups: int):
     """
     Computes the "whitening metric", a value which will be 1.0 if all the eigenvalues of
     of the centered feature covariance are the same within each group's covariance matrix
@@ -946,25 +992,22 @@ def _whitening_metric(x: Tensor,
     # the following expression is what we'd get if we took the matrix product
     # of each covariance and measured the mean of its trace, i.e.
     # the same as _diag(torch.matmul(x_covar, x_covar)).mean().
-    x_covarsq_mean_diag = (x_covar ** 2).sum() / (num_groups * channels_per_group)
+    x_covarsq_mean_diag = (x_covar**2).sum() / (num_groups * channels_per_group)
     # this metric will be >= 1.0; the larger it is, the less 'white' the data was.
-    metric = x_covarsq_mean_diag / (x_covar_mean_diag ** 2 + 1.0e-20)
+    metric = x_covarsq_mean_diag / (x_covar_mean_diag**2 + 1.0e-20)
     return metric
 
 
 class WhiteningPenaltyFunction(torch.autograd.Function):
     @staticmethod
-    def forward(ctx,
-                x: Tensor,
-                module: nn.Module) -> Tensor:
+    def forward(ctx, x: Tensor, module: nn.Module) -> Tensor:
         ctx.save_for_backward(x)
         ctx.module = module
         return x
 
     @staticmethod
-    def backward(ctx,
-                 x_grad: Tensor):
-        x_orig, = ctx.saved_tensors
+    def backward(ctx, x_grad: Tensor):
+        (x_orig,) = ctx.saved_tensors
         w = ctx.module
 
         try:
@@ -976,8 +1019,10 @@ class WhiteningPenaltyFunction(torch.autograd.Function):
                     metric = _whitening_metric(x_detached, w.num_groups)
 
                     if random.random() < 0.005 or __name__ == "__main__":
-                        logging.info(f"Whitening: name={w.name}, num_groups={w.num_groups}, num_channels={x_orig.shape[-1]}, "
-                                     f"metric={metric.item():.2f} vs. limit={float(w.whitening_limit)}")
+                        logging.info(
+                            f"Whitening: name={w.name}, num_groups={w.num_groups}, num_channels={x_orig.shape[-1]}, "
+                            f"metric={metric.item():.2f} vs. limit={float(w.whitening_limit)}"
+                        )
 
                     if metric < float(w.whitening_limit):
                         w.prob = w.min_prob
@@ -986,22 +1031,27 @@ class WhiteningPenaltyFunction(torch.autograd.Function):
                         w.prob = w.max_prob
                         metric.backward()
                         penalty_grad = x_detached.grad
-                        scale = w.grad_scale * (x_grad.to(torch.float32).norm() /
-                                                (penalty_grad.norm() + 1.0e-20))
+                        scale = w.grad_scale * (
+                            x_grad.to(torch.float32).norm()
+                            / (penalty_grad.norm() + 1.0e-20)
+                        )
                         penalty_grad = penalty_grad * scale
                         return x_grad + penalty_grad.to(x_grad.dtype), None
         except Exception as e:
-            logging.info(f"Caught exception in Whiten backward: {e}, size={list(x_grad.shape)}, will continue.")
+            logging.info(
+                f"Caught exception in Whiten backward: {e}, size={list(x_grad.shape)}, will continue."
+            )
         return x_grad, None
 
 
 class Whiten(nn.Module):
     def __init__(
-            self,
-            num_groups: int,
-            whitening_limit: FloatLike,
-            prob: Union[float, Tuple[float,float]],
-            grad_scale: FloatLike):
+        self,
+        num_groups: int,
+        whitening_limit: FloatLike,
+        prob: Union[float, Tuple[float, float]],
+        grad_scale: FloatLike,
+    ):
         """
         Args:
           num_groups: the number of groups to divide the channel dim into before
@@ -1033,10 +1083,9 @@ class Whiten(nn.Module):
         (self.min_prob, self.max_prob) = prob
         assert 0 < self.min_prob <= self.max_prob <= 1
         self.prob = self.max_prob
-        self.name = None # will be set in training loop
+        self.name = None  # will be set in training loop
 
-    def forward(self,
-                x: Tensor) -> Tensor:
+    def forward(self, x: Tensor) -> Tensor:
         """
         In the forward pass, this function just returns the input unmodified.
         In the backward pass, it will modify the gradients to ensure that the
@@ -1071,9 +1120,11 @@ class WithLoss(torch.autograd.Function):
 
     @staticmethod
     def backward(ctx, ans_grad: Tensor):
-        return ans_grad, torch.ones(ctx.y_shape,
-                                    dtype=ans_grad.dtype,
-                                    device=ans_grad.device), None
+        return (
+            ans_grad,
+            torch.ones(ctx.y_shape, dtype=ans_grad.dtype, device=ans_grad.device),
+            None,
+        )
 
 
 def with_loss(x, y, name):
@@ -1118,20 +1169,21 @@ class LimitParamValue(torch.autograd.Function):
 
     @staticmethod
     def backward(ctx, x_grad: Tensor):
-        x, = ctx.saved_tensors
+        (x,) = ctx.saved_tensors
         # where x < ctx.min, ensure all grads are negative (this will tend to make
         # x more positive).
-        x_grad = x_grad * torch.where(torch.logical_and(x_grad > 0, x < ctx.min), -1.0, 1.0)
+        x_grad = x_grad * torch.where(
+            torch.logical_and(x_grad > 0, x < ctx.min), -1.0, 1.0
+        )
         # where x > ctx.max, ensure all grads are positive (this will tend to make
         # x more negative).
         x_grad *= torch.where(torch.logical_and(x_grad < 0, x > ctx.max), -1.0, 1.0)
         return x_grad, None, None
 
 
-def limit_param_value(x: Tensor,
-                      min: float, max: float,
-                      prob: float = 0.6,
-                      training: bool = True):
+def limit_param_value(
+    x: Tensor, min: float, max: float, prob: float = 0.6, training: bool = True
+):
     # You apply this to (typically) an nn.Parameter during training to ensure that its
     # (elements mostly) stays within a supplied range.  This is done by modifying the
     # gradients in backprop.
@@ -1187,7 +1239,7 @@ class DoubleSwishFunction(torch.autograd.Function):
         y = x * s
 
         if requires_grad:
-            deriv = (y * (1 - s) + s)
+            deriv = y * (1 - s) + s
 
             # notes on derivative of x * sigmoid(x - 1):
             # https://www.wolframalpha.com/input?i=d%2Fdx+%28x+*+sigmoid%28x-1%29%29
@@ -1197,7 +1249,9 @@ class DoubleSwishFunction(torch.autograd.Function):
             # floors), should be expectation-preserving.
             floor = -0.044
             ceil = 1.2
-            d_scaled = ((deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like(deriv))
+            d_scaled = (deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                deriv
+            )
             if __name__ == "__main__":
                 # for self-testing only.
                 assert d_scaled.min() >= 0.0
@@ -1210,12 +1264,12 @@ class DoubleSwishFunction(torch.autograd.Function):
 
     @staticmethod
     def backward(ctx, y_grad: Tensor) -> Tensor:
-        d, = ctx.saved_tensors
+        (d,) = ctx.saved_tensors
         # the same constants as used in forward pass.
         floor = -0.043637
         ceil = 1.2
 
-        d = (d * ((ceil - floor) / 255.0) + floor)
+        d = d * ((ceil - floor) / 255.0) + floor
         return y_grad * d
 
 
@@ -1239,9 +1293,7 @@ class Dropout2(nn.Module):
         self.p = p
 
     def forward(self, x: Tensor) -> Tensor:
-        return torch.nn.functional.dropout(x,
-                                           p=float(self.p),
-                                           training=self.training)
+        return torch.nn.functional.dropout(x, p=float(self.p), training=self.training)
 
 
 class MulForDropout3(torch.autograd.Function):
@@ -1259,7 +1311,7 @@ class MulForDropout3(torch.autograd.Function):
     @staticmethod
     @custom_bwd
     def backward(ctx, ans_grad):
-        ans, = ctx.saved_tensors
+        (ans,) = ctx.saved_tensors
         x_grad = ctx.alpha * ans_grad * (ans != 0)
         return x_grad, None, None
 
@@ -1286,7 +1338,7 @@ class Dropout3(nn.Module):
 
 class SwooshLFunction(torch.autograd.Function):
     """
-      swoosh_l(x) =  log(1 + exp(x-4)) - 0.08*x - 0.035
+    swoosh_l(x) =  log(1 + exp(x-4)) - 0.08*x - 0.035
     """
 
     @staticmethod
@@ -1308,13 +1360,15 @@ class SwooshLFunction(torch.autograd.Function):
                 if not requires_grad:
                     return y
 
-                y.backward(gradient = torch.ones_like(y))
+                y.backward(gradient=torch.ones_like(y))
 
                 grad = x.grad
                 floor = coeff
                 ceil = 1.0 + coeff + 0.005
 
-                d_scaled = ((grad - floor) * (255.0 / (ceil - floor)) + torch.rand_like(grad))
+                d_scaled = (grad - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                    grad
+                )
                 if __name__ == "__main__":
                     # for self-testing only.
                     assert d_scaled.min() >= 0.0
@@ -1328,20 +1382,19 @@ class SwooshLFunction(torch.autograd.Function):
 
     @staticmethod
     def backward(ctx, y_grad: Tensor) -> Tensor:
-        d, = ctx.saved_tensors
+        (d,) = ctx.saved_tensors
         # the same constants as used in forward pass.
 
         coeff = -0.08
         floor = coeff
         ceil = 1.0 + coeff + 0.005
-        d = (d * ((ceil - floor) / 255.0) + floor)
-        return (y_grad * d)
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
 
 
 class SwooshL(torch.nn.Module):
     def forward(self, x: Tensor) -> Tensor:
-        """Return Swoosh-L activation.
-        """
+        """Return Swoosh-L activation."""
         if torch.jit.is_scripting() or torch.jit.is_tracing():
             zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
             return logaddexp(zero, x - 4.0) - 0.08 * x - 0.035
@@ -1351,19 +1404,19 @@ class SwooshL(torch.nn.Module):
             return k2.swoosh_l(x)
         # return SwooshLFunction.apply(x)
 
+
 class SwooshLOnnx(torch.nn.Module):
     def forward(self, x: Tensor) -> Tensor:
-        """Return Swoosh-L activation.
-        """
+        """Return Swoosh-L activation."""
         zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
         return logaddexp_onnx(zero, x - 4.0) - 0.08 * x - 0.035
 
 
 class SwooshRFunction(torch.autograd.Function):
     """
-      swoosh_r(x) =  log(1 + exp(x-1)) - 0.08*x - 0.313261687
+     swoosh_r(x) =  log(1 + exp(x-1)) - 0.08*x - 0.313261687
 
-     derivatives are between -0.08 and 0.92.
+    derivatives are between -0.08 and 0.92.
     """
 
     @staticmethod
@@ -1379,17 +1432,19 @@ class SwooshRFunction(torch.autograd.Function):
             with torch.enable_grad():
                 x = x.detach()
                 x.requires_grad = True
-                y = torch.logaddexp(zero, x - 1.) - 0.08 * x - 0.313261687
+                y = torch.logaddexp(zero, x - 1.0) - 0.08 * x - 0.313261687
 
                 if not requires_grad:
                     return y
-                y.backward(gradient = torch.ones_like(y))
+                y.backward(gradient=torch.ones_like(y))
 
                 grad = x.grad
                 floor = -0.08
                 ceil = 0.925
 
-                d_scaled = ((grad - floor) * (255.0 / (ceil - floor)) + torch.rand_like(grad))
+                d_scaled = (grad - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                    grad
+                )
                 if __name__ == "__main__":
                     # for self-testing only.
                     assert d_scaled.min() >= 0.0
@@ -1403,33 +1458,32 @@ class SwooshRFunction(torch.autograd.Function):
 
     @staticmethod
     def backward(ctx, y_grad: Tensor) -> Tensor:
-        d, = ctx.saved_tensors
+        (d,) = ctx.saved_tensors
         # the same constants as used in forward pass.
         floor = -0.08
         ceil = 0.925
-        d = (d * ((ceil - floor) / 255.0) + floor)
-        return (y_grad * d)
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
 
 
 class SwooshR(torch.nn.Module):
     def forward(self, x: Tensor) -> Tensor:
-        """Return Swoosh-R activation.
-        """
+        """Return Swoosh-R activation."""
         if torch.jit.is_scripting() or torch.jit.is_tracing():
             zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
-            return logaddexp(zero, x - 1.) - 0.08 * x - 0.313261687
+            return logaddexp(zero, x - 1.0) - 0.08 * x - 0.313261687
         if not x.requires_grad:
             return k2.swoosh_r_forward(x)
         else:
             return k2.swoosh_r(x)
         # return SwooshRFunction.apply(x)
 
+
 class SwooshROnnx(torch.nn.Module):
     def forward(self, x: Tensor) -> Tensor:
-        """Return Swoosh-R activation.
-        """
+        """Return Swoosh-R activation."""
         zero = torch.tensor(0.0, dtype=x.dtype, device=x.device)
-        return logaddexp_onnx(zero, x - 1.) - 0.08 * x - 0.313261687
+        return logaddexp_onnx(zero, x - 1.0) - 0.08 * x - 0.313261687
 
 
 # simple version of SwooshL that does not redefine the backprop, used in
@@ -1437,7 +1491,7 @@ class SwooshROnnx(torch.nn.Module):
 def SwooshLForward(x: Tensor):
     x_offset = x - 4.0
     log_sum = (1.0 + x_offset.exp()).log().to(x.dtype)
-    log_sum = torch.where(log_sum == float('inf'), x_offset, log_sum)
+    log_sum = torch.where(log_sum == float("inf"), x_offset, log_sum)
     return log_sum - 0.08 * x - 0.035
 
 
@@ -1446,28 +1500,30 @@ def SwooshLForward(x: Tensor):
 def SwooshRForward(x: Tensor):
     x_offset = x - 1.0
     log_sum = (1.0 + x_offset.exp()).log().to(x.dtype)
-    log_sum = torch.where(log_sum == float('inf'), x_offset, log_sum)
+    log_sum = torch.where(log_sum == float("inf"), x_offset, log_sum)
     return log_sum - 0.08 * x - 0.313261687
 
 
 class ActivationDropoutAndLinearFunction(torch.autograd.Function):
     @staticmethod
     @custom_fwd
-    def forward(ctx,
-                x: Tensor,
-                weight: Tensor,
-                bias: Optional[Tensor],
-                activation: str,
-                dropout_p: float,
-                dropout_shared_dim: Optional[int]):
+    def forward(
+        ctx,
+        x: Tensor,
+        weight: Tensor,
+        bias: Optional[Tensor],
+        activation: str,
+        dropout_p: float,
+        dropout_shared_dim: Optional[int],
+    ):
         if dropout_p != 0.0:
             dropout_shape = list(x.shape)
             if dropout_shared_dim is not None:
                 dropout_shape[dropout_shared_dim] = 1
             # else it won't be very memory efficient.
-            dropout_mask = ((1.0 / (1.0 - dropout_p)) *
-                            (torch.rand(*dropout_shape,
-                                        device=x.device, dtype=x.dtype) > dropout_p))
+            dropout_mask = (1.0 / (1.0 - dropout_p)) * (
+                torch.rand(*dropout_shape, device=x.device, dtype=x.dtype) > dropout_p
+            )
         else:
             dropout_mask = None
 
@@ -1476,8 +1532,8 @@ class ActivationDropoutAndLinearFunction(torch.autograd.Function):
         ctx.activation = activation
 
         forward_activation_dict = {
-            'SwooshL': k2.swoosh_l_forward,
-            'SwooshR': k2.swoosh_r_forward
+            "SwooshL": k2.swoosh_l_forward,
+            "SwooshR": k2.swoosh_r_forward,
         }
         # it will raise a KeyError if this fails.  This will be an error.  We let it
         # propagate to the user.
@@ -1495,8 +1551,8 @@ class ActivationDropoutAndLinearFunction(torch.autograd.Function):
         (x, weight, bias, dropout_mask) = saved
 
         forward_and_deriv_activation_dict = {
-            'SwooshL': k2.swoosh_l_forward_and_deriv,
-            'SwooshR': k2.swoosh_r_forward_and_deriv
+            "SwooshL": k2.swoosh_l_forward_and_deriv,
+            "SwooshR": k2.swoosh_r_forward_and_deriv,
         }
         # the following lines a KeyError if the activation is unrecognized.
         # This will be an error.  We let it propagate to the user.
@@ -1511,8 +1567,7 @@ class ActivationDropoutAndLinearFunction(torch.autograd.Function):
 
         in_channels = y.shape[-1]
         g = ans_grad.reshape(-1, out_channels)
-        weight_deriv = torch.matmul(g.t(),
-                                    y.reshape(-1, in_channels))
+        weight_deriv = torch.matmul(g.t(), y.reshape(-1, in_channels))
         y_deriv = torch.matmul(ans_grad, weight)
         bias_deriv = None if bias is None else g.sum(dim=0)
         x_deriv = y_deriv * func_deriv
@@ -1525,71 +1580,76 @@ class ActivationDropoutAndLinearFunction(torch.autograd.Function):
 
 class ActivationDropoutAndLinear(torch.nn.Module):
     """
-    This merges an activation function followed by dropout and then a nn.Linear module;
-    it does so in a memory efficient way so that it only stores the input to the whole
-    module.  If activation == SwooshL and dropout_shared_dim != None, this will be
-    equivalent to:
-      nn.Sequential(SwooshL(),
-                    Dropout3(dropout_p, shared_dim=dropout_shared_dim),
-                    ScaledLinear(in_channels, out_channels, bias=bias,
-                                 initial_scale=initial_scale))
-   If dropout_shared_dim is None, the dropout would be equivalent to
-   Dropout2(dropout_p).  Note: Dropout3 will be more memory efficient as the dropout
-   mask is smaller.
+     This merges an activation function followed by dropout and then a nn.Linear module;
+     it does so in a memory efficient way so that it only stores the input to the whole
+     module.  If activation == SwooshL and dropout_shared_dim != None, this will be
+     equivalent to:
+       nn.Sequential(SwooshL(),
+                     Dropout3(dropout_p, shared_dim=dropout_shared_dim),
+                     ScaledLinear(in_channels, out_channels, bias=bias,
+                                  initial_scale=initial_scale))
+    If dropout_shared_dim is None, the dropout would be equivalent to
+    Dropout2(dropout_p).  Note: Dropout3 will be more memory efficient as the dropout
+    mask is smaller.
 
-    Args:
-       in_channels: number of input channels, e.g. 256
-       out_channels: number of output channels, e.g. 256
-       bias: if true, have a bias
-       activation: the activation function, for now just support SwooshL.
-       dropout_p: the dropout probability or schedule (happens after nonlinearity).
-       dropout_shared_dim: the dimension, if any, across which the dropout mask is
-            shared (e.g. the time dimension).  If None, this may be less memory
-            efficient if there are modules before this one that cache the input
-            for their backprop (e.g. Balancer or Whiten).
+     Args:
+        in_channels: number of input channels, e.g. 256
+        out_channels: number of output channels, e.g. 256
+        bias: if true, have a bias
+        activation: the activation function, for now just support SwooshL.
+        dropout_p: the dropout probability or schedule (happens after nonlinearity).
+        dropout_shared_dim: the dimension, if any, across which the dropout mask is
+             shared (e.g. the time dimension).  If None, this may be less memory
+             efficient if there are modules before this one that cache the input
+             for their backprop (e.g. Balancer or Whiten).
     """
-    def __init__(self,
-                 in_channels: int,
-                 out_channels: int,
-                 bias: bool = True,
-                 activation: str = 'SwooshL',
-                 dropout_p: FloatLike = 0.0,
-                 dropout_shared_dim: Optional[int] = -1,
-                 initial_scale: float = 1.0):
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        bias: bool = True,
+        activation: str = "SwooshL",
+        dropout_p: FloatLike = 0.0,
+        dropout_shared_dim: Optional[int] = -1,
+        initial_scale: float = 1.0,
+    ):
         super().__init__()
         # create a temporary module of nn.Linear that we'll steal the
         # weights and bias from
-        l = ScaledLinear(in_channels, out_channels,
-                         bias=bias,
-                         initial_scale=initial_scale)
+        l = ScaledLinear(
+            in_channels, out_channels, bias=bias, initial_scale=initial_scale
+        )
 
         self.weight = l.weight
         # register_parameter properly handles making it a parameter when l.bias
         # is None. I think there is some reason for doing it this way rather
         # than just setting it to None but I don't know what it is, maybe
         # something to do with exporting the module..
-        self.register_parameter('bias', l.bias)
+        self.register_parameter("bias", l.bias)
 
         self.activation = activation
         self.dropout_p = dropout_p
         self.dropout_shared_dim = dropout_shared_dim
 
-    def forward(self,
-                x: Tensor):
+    def forward(self, x: Tensor):
         if torch.jit.is_scripting() or torch.jit.is_tracing():
-            if self.activation == 'SwooshL':
+            if self.activation == "SwooshL":
                 x = SwooshLForward(x)
             elif self.activation == "SwooshR":
                 x = SwooshRForward(x)
             else:
                 assert False, self.activation
-            return torch.nn.functional.linear(x,
-                                              self.weight,
-                                              self.bias)
+            return torch.nn.functional.linear(x, self.weight, self.bias)
 
         return ActivationDropoutAndLinearFunction.apply(
-            x, self.weight, self.bias, self.activation,
-            float(self.dropout_p), self.dropout_shared_dim)
+            x,
+            self.weight,
+            self.bias,
+            self.activation,
+            float(self.dropout_p),
+            self.dropout_shared_dim,
+        )
 
 
 def convert_num_channels(x: Tensor, num_channels: int) -> Tensor:
@@ -1612,10 +1672,9 @@ def _test_whiten():
 
         x.requires_grad = True
 
-        m = Whiten(1,  # num_groups
-                   5.0,  # whitening_limit,
-                   prob=1.0,
-                   grad_scale=0.1)  # grad_scale
+        m = Whiten(
+            1, 5.0, prob=1.0, grad_scale=0.1  # num_groups  # whitening_limit,
+        )  # grad_scale
 
         for _ in range(4):
             y = m(x)
@@ -1656,9 +1715,7 @@ def _test_balancer_sign():
 def _test_balancer_magnitude():
     magnitudes = torch.arange(0, 1, 0.01)
     N = 1000
-    x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(
-        -1
-    )
+    x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(-1)
     x = x.detach()
     x.requires_grad = True
     m = Balancer(
@@ -1685,7 +1742,7 @@ def _test_double_swish_deriv():
     x.requires_grad = True
     m = DoubleSwish()
 
-    tol = ((1.2-(-0.043637))/255.0)
+    tol = (1.2 - (-0.043637)) / 255.0
     torch.autograd.gradcheck(m, x, atol=tol)
 
     # for self-test.
@@ -1699,7 +1756,7 @@ def _test_swooshl_deriv():
     x.requires_grad = True
     m = SwooshL()
 
-    tol = (1.0 / 255.0)
+    tol = 1.0 / 255.0
     torch.autograd.gradcheck(m, x, atol=tol, eps=0.01)
 
     # for self-test.
@@ -1713,7 +1770,7 @@ def _test_swooshr_deriv():
     x.requires_grad = True
     m = SwooshR()
 
-    tol = (1.0 / 255.0)
+    tol = 1.0 / 255.0
     torch.autograd.gradcheck(m, x, atol=tol, eps=0.01)
 
     # for self-test.
@@ -1727,24 +1784,24 @@ def _test_softmax():
     b = a.clone()
     a.requires_grad = True
     b.requires_grad = True
-    a.softmax(dim=1)[:,0].sum().backward()
+    a.softmax(dim=1)[:, 0].sum().backward()
     print("a grad = ", a.grad)
-    softmax(b, dim=1)[:,0].sum().backward()
+    softmax(b, dim=1)[:, 0].sum().backward()
     print("b grad = ", b.grad)
     assert torch.allclose(a.grad, b.grad)
 
 
 def _test_piecewise_linear():
-    p = PiecewiseLinear( (0, 10.0) )
+    p = PiecewiseLinear((0, 10.0))
     for x in [-100, 0, 100]:
         assert p(x) == 10.0
-    p = PiecewiseLinear( (0, 10.0), (1, 0.0) )
-    for x, y in [ (-100, 10.0), (0, 10.0), (0.5, 5.0), (1, 0.0), (2, 0.0) ]:
+    p = PiecewiseLinear((0, 10.0), (1, 0.0))
+    for x, y in [(-100, 10.0), (0, 10.0), (0.5, 5.0), (1, 0.0), (2, 0.0)]:
         print("x, y = ", x, y)
         assert p(x) == y, (x, p(x), y)
 
     q = PiecewiseLinear((0.5, 15.0), (0.6, 1.0))
-    x_vals = [ -1.0, 0.0, 0.1, 0.2, 0.5, 0.6, 0.7, 0.9, 1.0, 2.0 ]
+    x_vals = [-1.0, 0.0, 0.1, 0.2, 0.5, 0.6, 0.7, 0.9, 1.0, 2.0]
     pq = p.max(q)
     for x in x_vals:
         y1 = max(p(x), q(x))
@@ -1757,7 +1814,7 @@ def _test_piecewise_linear():
         assert abs(y1 - y2) < 0.001
     pq = p + q
     for x in x_vals:
-        y1 = p(x) +  q(x)
+        y1 = p(x) + q(x)
         y2 = pq(x)
         assert abs(y1 - y2) < 0.001
 
@@ -1772,15 +1829,22 @@ def _test_activation_dropout_and_linear():
         # swoosh_l an swoosh_r inside SwooshL() and SwooshR(), and they call randn()
         # internally, messing up the random state.
         for dropout_p in [0.0]:
-            for activation in ['SwooshL', 'SwooshR']:
-                m1 =  nn.Sequential(SwooshL() if activation == 'SwooshL' else SwooshR(),
-                                    Dropout3(p=dropout_p, shared_dim=-1),
-                                    ScaledLinear(in_channels, out_channels, bias=bias,
-                                                 initial_scale=0.5))
-                m2 = ActivationDropoutAndLinear(in_channels, out_channels,
-                                                bias=bias, initial_scale=0.5,
-                                                activation=activation,
-                                                dropout_p=dropout_p)
+            for activation in ["SwooshL", "SwooshR"]:
+                m1 = nn.Sequential(
+                    SwooshL() if activation == "SwooshL" else SwooshR(),
+                    Dropout3(p=dropout_p, shared_dim=-1),
+                    ScaledLinear(
+                        in_channels, out_channels, bias=bias, initial_scale=0.5
+                    ),
+                )
+                m2 = ActivationDropoutAndLinear(
+                    in_channels,
+                    out_channels,
+                    bias=bias,
+                    initial_scale=0.5,
+                    activation=activation,
+                    dropout_p=dropout_p,
+                )
                 with torch.no_grad():
                     m2.weight[:] = m1[2].weight
                     if bias:
@@ -1790,9 +1854,9 @@ def _test_activation_dropout_and_linear():
                 x1.requires_grad = True
 
                 # TEMP.
-                assert torch.allclose(SwooshRFunction.apply(x1),
-                                      SwooshRForward(x1),
-                                      atol=1.0e-03)
+                assert torch.allclose(
+                    SwooshRFunction.apply(x1), SwooshRForward(x1), atol=1.0e-03
+                )
 
                 x2 = x1.clone().detach()
                 x2.requires_grad = True
@@ -1805,21 +1869,24 @@ def _test_activation_dropout_and_linear():
                 y2 = m2(x2)
                 y2.backward(gradient=y_grad)
 
-                print(f"bias = {bias}, dropout_p = {dropout_p}, activation = {activation}")
+                print(
+                    f"bias = {bias}, dropout_p = {dropout_p}, activation = {activation}"
+                )
                 print("y1 = ", y1)
                 print("y2 = ", y2)
                 assert torch.allclose(y1, y2, atol=0.02)
-                assert torch.allclose(m1[2].weight.grad, m2.weight.grad,
-                                      atol=1.0e-05)
+                assert torch.allclose(m1[2].weight.grad, m2.weight.grad, atol=1.0e-05)
                 if bias:
-                    assert torch.allclose(m1[2].bias.grad, m2.bias.grad,
-                                          atol=1.0e-05)
+                    assert torch.allclose(m1[2].bias.grad, m2.bias.grad, atol=1.0e-05)
                 print("x1.grad = ", x1.grad)
                 print("x2.grad = ", x2.grad)
 
                 def isclose(a, b):
                     # return true if cosine similarity is > 0.9.
-                    return (a * b).sum() > 0.9 * ((a**2).sum() * (b**2).sum()).sqrt()
+                    return (a * b).sum() > 0.9 * (
+                        (a**2).sum() * (b**2).sum()
+                    ).sqrt()
+
                 # the SwooshL() implementation has a noisy gradient due to 1-byte
                 # storage of it.
                 assert isclose(x1.grad, x2.grad)
diff --git a/egs/librispeech/ASR/zipformer/streaming_decode.py b/egs/librispeech/ASR/zipformer/streaming_decode.py
index 44ff392a3..904caf8af 100755
--- a/egs/librispeech/ASR/zipformer/streaming_decode.py
+++ b/egs/librispeech/ASR/zipformer/streaming_decode.py
@@ -282,9 +282,7 @@ def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]:
     )
     batch_states.append(cached_embed_left_pad)
 
-    processed_lens = torch.cat(
-        [state_list[i][-1] for i in range(batch_size)], dim=0
-    )
+    processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0)
     batch_states.append(processed_lens)
 
     return batch_states
@@ -322,9 +320,7 @@ def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]:
     for layer in range(tot_num_layers):
         layer_offset = layer * 6
         # cached_key: (left_context_len, batch_size, key_dim)
-        cached_key_list = batch_states[layer_offset].chunk(
-            chunks=batch_size, dim=1
-        )
+        cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1)
         # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim)
         cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk(
             chunks=batch_size, dim=1
@@ -355,9 +351,7 @@ def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]:
                 cached_conv2_list[i],
             ]
 
-    cached_embed_left_pad_list = batch_states[-2].chunk(
-        chunks=batch_size, dim=0
-    )
+    cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0)
     for i in range(batch_size):
         state_list[i].append(cached_embed_left_pad_list[i])
 
@@ -380,11 +374,7 @@ def streaming_forward(
     Returns encoder outputs, output lengths, and updated states.
     """
     cached_embed_left_pad = states[-2]
-    (
-        x,
-        x_lens,
-        new_cached_embed_left_pad,
-    ) = model.encoder_embed.streaming_forward(
+    (x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward(
         x=features,
         x_lens=feature_lens,
         cached_left_pad=cached_embed_left_pad,
@@ -404,9 +394,7 @@ def streaming_forward(
     new_processed_lens = processed_lens + x_lens
 
     # (batch, left_context_size + chunk_size)
-    src_key_padding_mask = torch.cat(
-        [processed_mask, src_key_padding_mask], dim=1
-    )
+    src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1)
 
     x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
     encoder_states = states[:-2]
@@ -494,9 +482,7 @@ def decode_one_chunk(
     encoder_out = model.joiner.encoder_proj(encoder_out)
 
     if params.decoding_method == "greedy_search":
-        greedy_search(
-            model=model, encoder_out=encoder_out, streams=decode_streams
-        )
+        greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams)
     elif params.decoding_method == "fast_beam_search":
         processed_lens = torch.tensor(processed_lens, device=device)
         processed_lens = processed_lens + encoder_out_lens
@@ -517,9 +503,7 @@ def decode_one_chunk(
             num_active_paths=params.num_active_paths,
         )
     else:
-        raise ValueError(
-            f"Unsupported decoding method: {params.decoding_method}"
-        )
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
 
     states = unstack_states(new_states)
 
@@ -577,9 +561,7 @@ def decode_dataset(
     decode_streams = []
     for num, cut in enumerate(cuts):
         # each utterance has a DecodeStream.
-        initial_states = get_init_states(
-            model=model, batch_size=1, device=device
-        )
+        initial_states = get_init_states(model=model, batch_size=1, device=device)
         decode_stream = DecodeStream(
             params=params,
             cut_id=cut.id,
@@ -649,9 +631,7 @@ def decode_dataset(
     elif params.decoding_method == "modified_beam_search":
         key = f"num_active_paths_{params.num_active_paths}"
     else:
-        raise ValueError(
-            f"Unsupported decoding method: {params.decoding_method}"
-        )
+        raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
     return {key: decode_results}
 
 
@@ -684,8 +664,7 @@ def save_results(
 
     test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
     errs_info = (
-        params.res_dir
-        / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
     )
     with open(errs_info, "w") as f:
         print("settings\tWER", file=f)
@@ -718,9 +697,7 @@ def main():
         params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
 
     assert params.causal, params.causal
-    assert (
-        "," not in params.chunk_size
-    ), "chunk_size should be one value in decoding."
+    assert "," not in params.chunk_size, "chunk_size should be one value in decoding."
     assert (
         "," not in params.left_context_frames
     ), "left_context_frames should be one value in decoding."
@@ -760,9 +737,9 @@ def main():
 
     if not params.use_averaged_model:
         if params.iter > 0:
-            filenames = find_checkpoints(
-                params.exp_dir, iteration=-params.iter
-            )[: params.avg]
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
             if len(filenames) == 0:
                 raise ValueError(
                     f"No checkpoints found for"
@@ -789,9 +766,9 @@ def main():
             model.load_state_dict(average_checkpoints(filenames, device=device))
     else:
         if params.iter > 0:
-            filenames = find_checkpoints(
-                params.exp_dir, iteration=-params.iter
-            )[: params.avg + 1]
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
             if len(filenames) == 0:
                 raise ValueError(
                     f"No checkpoints found for"
diff --git a/egs/librispeech/ASR/zipformer/subsampling.py b/egs/librispeech/ASR/zipformer/subsampling.py
index 6532ddccb..d16d87bac 100644
--- a/egs/librispeech/ASR/zipformer/subsampling.py
+++ b/egs/librispeech/ASR/zipformer/subsampling.py
@@ -107,9 +107,7 @@ class ConvNeXt(nn.Module):
         if layerdrop_rate != 0.0:
             batch_size = x.shape[0]
             mask = (
-                torch.rand(
-                    (batch_size, 1, 1, 1), dtype=x.dtype, device=x.device
-                )
+                torch.rand((batch_size, 1, 1, 1), dtype=x.dtype, device=x.device)
                 > layerdrop_rate
             )
         else:
@@ -278,9 +276,7 @@ class Conv2dSubsampling(nn.Module):
         # many copies of this extra gradient term.
         self.out_whiten = Whiten(
             num_groups=1,
-            whitening_limit=ScheduledFloat(
-                (0.0, 4.0), (20000.0, 8.0), default=4.0
-            ),
+            whitening_limit=ScheduledFloat((0.0, 4.0), (20000.0, 8.0), default=4.0),
             prob=(0.025, 0.25),
             grad_scale=0.02,
         )
@@ -331,7 +327,7 @@ class Conv2dSubsampling(nn.Module):
             with warnings.catch_warnings():
                 warnings.simplefilter("ignore")
                 x_lens = (x_lens - 7) // 2
-        assert x.size(1) == x_lens.max().item() , (x.size(1), x_lens.max())
+        assert x.size(1) == x_lens.max().item(), (x.size(1), x_lens.max())
 
         return x, x_lens
 
@@ -403,8 +399,8 @@ class Conv2dSubsampling(nn.Module):
         left_pad = self.convnext.padding[0]
         freq = self.out_width
         channels = self.layer3_channels
-        cached_embed_left_pad = torch.zeros(
-            batch_size, channels, left_pad, freq
-        ).to(device)
+        cached_embed_left_pad = torch.zeros(batch_size, channels, left_pad, freq).to(
+            device
+        )
 
         return cached_embed_left_pad
diff --git a/egs/librispeech/ASR/zipformer/train.py b/egs/librispeech/ASR/zipformer/train.py
index bc3e9c1ba..7009f3346 100755
--- a/egs/librispeech/ASR/zipformer/train.py
+++ b/egs/librispeech/ASR/zipformer/train.py
@@ -604,11 +604,11 @@ def get_joiner_model(params: AttributeDict) -> nn.Module:
 
 
 def get_model(params: AttributeDict) -> nn.Module:
-    assert (
-        params.use_transducer or params.use_ctc
-    ), (f"At least one of them should be True, "
+    assert params.use_transducer or params.use_ctc, (
+        f"At least one of them should be True, "
         f"but got params.use_transducer={params.use_transducer}, "
-        f"params.use_ctc={params.use_ctc}")
+        f"params.use_ctc={params.use_ctc}"
+    )
 
     encoder_embed = get_encoder_embed(params)
     encoder = get_encoder_model(params)
@@ -808,17 +808,16 @@ def compute_loss(
             # take down the scale on the simple loss from 1.0 at the start
             # to params.simple_loss scale by warm_step.
             simple_loss_scale = (
-                s if batch_idx_train >= warm_step
+                s
+                if batch_idx_train >= warm_step
                 else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
             )
             pruned_loss_scale = (
-                1.0 if batch_idx_train >= warm_step
+                1.0
+                if batch_idx_train >= warm_step
                 else 0.1 + 0.9 * (batch_idx_train / warm_step)
             )
-            loss += (
-                simple_loss_scale * simple_loss
-                + pruned_loss_scale * pruned_loss
-            )
+            loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss
 
         if params.use_ctc:
             loss += params.ctc_loss_scale * ctc_loss
@@ -1166,7 +1165,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/ASR/zipformer/zipformer.py b/egs/librispeech/ASR/zipformer/zipformer.py
index 1a174b315..61ae378d8 100644
--- a/egs/librispeech/ASR/zipformer/zipformer.py
+++ b/egs/librispeech/ASR/zipformer/zipformer.py
@@ -17,28 +17,33 @@
 # limitations under the License.
 
 import copy
+import logging
 import math
+import random
 import warnings
 from typing import List, Optional, Tuple, Union
-import logging
+
 import torch
-import random
 from encoder_interface import EncoderInterface
 from scaling import (
+    Identity,  # more friendly to backward hooks than nn.Identity(), for diagnostic reasons.
+)
+from scaling import (
+    ScaledLinear,  # not as in other dirs.. just scales down initial parameter values.
+)
+from scaling import (
+    ActivationDropoutAndLinear,
     Balancer,
     BiasNorm,
-    Dropout2,
     ChunkCausalDepthwiseConv1d,
-    ActivationDropoutAndLinear,
-    ScaledLinear,  # not as in other dirs.. just scales down initial parameter values.
+    Dropout2,
+    FloatLike,
+    ScheduledFloat,
     Whiten,
-    Identity,  # more friendly to backward hooks than nn.Identity(), for diagnostic reasons.
+    convert_num_channels,
+    limit_param_value,
     penalize_abs_values_gt,
     softmax,
-    ScheduledFloat,
-    FloatLike,
-    limit_param_value,
-    convert_num_channels,
 )
 from torch import Tensor, nn
 
@@ -2098,7 +2103,7 @@ class NonlinAttention(nn.Module):
         (seq_len, batch_size, _) = x.shape
         hidden_channels = self.hidden_channels
 
-        s, x, y = x.chunk(3, dim=-1)
+        s, x, y = x.chunk(3, dim=2)
 
         # s will go through tanh.
 
@@ -2151,7 +2156,7 @@ class NonlinAttention(nn.Module):
         (seq_len, batch_size, _) = x.shape
         hidden_channels = self.hidden_channels
 
-        s, x, y = x.chunk(3, dim=-1)
+        s, x, y = x.chunk(3, dim=2)
 
         # s will go through tanh.
         s = self.tanh(s)
@@ -2308,7 +2313,7 @@ class ConvolutionModule(nn.Module):
 
         x = self.in_proj(x)  # (time, batch, 2*channels)
 
-        x, s = x.chunk(2, dim=-1)
+        x, s = x.chunk(2, dim=2)
         s = self.balancer1(s)
         s = self.sigmoid(s)
         x = self.activation1(x)  # identity.
diff --git a/egs/librispeech/ASR/zipformer_mmi/train.py b/egs/librispeech/ASR/zipformer_mmi/train.py
index c1b3ea3e0..4b50acdde 100755
--- a/egs/librispeech/ASR/zipformer_mmi/train.py
+++ b/egs/librispeech/ASR/zipformer_mmi/train.py
@@ -981,7 +981,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/librispeech/WSASR/README.md b/egs/librispeech/WSASR/README.md
new file mode 100644
index 000000000..3b8822fd2
--- /dev/null
+++ b/egs/librispeech/WSASR/README.md
@@ -0,0 +1,224 @@
+# Introduction
+
+This is a weakly supervised ASR recipe for the LibriSpeech (clean 100 hours) dataset. We train a
+conformer model using [Bypass Temporal Classification](https://arxiv.org/pdf/2306.01031.pdf) (BTC)/[Omni-temporal Classification](https://arxiv.org/pdf/2309.15796.pdf) (OTC) with transcripts with synthetic errors. In this README, we will describe
+the task and the BTC/OTC training process.
+
+Note that OTC is an extension of BTC and supports all BTC functions. Therefore, in the following, we only describe OTC.
+## Task
+We propose BTC/OTC to directly train an ASR system leveraging weak supervision, i.e., speech with non-verbatim transcripts. This is achieved by using a special token $\star$ to model uncertainties (i.e., substitution errors, insertion errors, and deletion errors) 
+within the WFST framework during training.
+
+
+<div style="display: flex;flex; justify-content: space-between">
+  <figure style="flex: 2; text-align: center; margin: 5px;">
+    <img src="figures/sub.png" alt="Image 1" width="25%" />
+
+  </figure>
+  <figure style="flex: 2; text-align: center; margin: 5px;">
+    <img src="figures/ins.png" alt="Image 2" width="25%" />
+
+  </figure>
+  <figure style="flex: 2; text-align: center;margin: 5px;">
+    <img src="figures/del.png" alt="Image 3" width="25%" />
+
+  </figure>
+</div>
+<figcaption> Examples of errors (substitution, insertion, and deletion) in the transcript. The grey box is the verbatim transcript and the red box is the inaccurate transcript. Inaccurate words are marked in bold.</figcaption> <br><br>
+
+
+We modify $G(\mathbf{y})$ by adding self-loop arcs into each state and bypass arcs into each arc. 
+  <p align="center">
+    <img src="figures/otc_g.png" alt="Image Alt Text" width="50%" />
+
+  </p>
+
+We incorporate the penalty strategy and apply different configurations for the self-loop arc and bypass arc. The penalties are set as
+
+$$\lambda_{1_{i}} = \beta_{1} * \tau_{1}^{i},\quad \lambda_{2_{i}} = \beta_{2} * \tau_{2}^{i}$$
+
+for the $i$-th training epoch. $\beta$ is the initial penalty that encourages the model to rely more on the given transcript at the start of training. 
+It decays exponentially by a factor of $\tau \in (0, 1)$, gradually encouraging the model to align speech with $\star$ when getting confused. 
+
+After composing the modified WFST $G_{\text{otc}}(\mathbf{y})$ with $L$ and $T$, the OTC training graph is shown in this figure:
+<figure style="text-align: center">
+  <img src="figures/otc_training_graph.drawio.png" alt="Image Alt Text" />
+  <figcaption>OTC training graph. The self-loop arcs and bypass arcs are highlighted in green and blue, respectively.</figcaption>
+</figure>
+
+The $\star$ is represented as the average probability of all non-blank tokens.
+  <p align="center">
+    <img src="figures/otc_emission.drawio.png" width="50%" />
+  </p>
+
+The weight of $\star$ is the log average probability of "a" and "b": $\log \frac{e^{-1.2} + e^{-2.3}}{2} = -1.6$ and $\log \frac{e^{-1.9} + e^{-0.5}}{2} = -1.0$ for 2 frames.
+
+## Description of the recipe
+### Preparation
+```
+# feature_type can be ssl or fbank
+feature_type=ssl
+feature_dir="data/${feature_type}"
+manifest_dir="${feature_dir}"
+lang_dir="data/lang"
+lm_dir="data/lm"
+exp_dir="conformer_ctc2/exp"
+otc_token="<star>"
+
+./prepare.sh \
+  --feature-type "${feature_type}" \
+  --feature-dir "${feature_dir}" \
+  --lang-dir "${lang_dir}" \
+  --lm-dir "${lm_dir}" \
+  --otc-token "${otc_token}" 
+```
+This script adds the 'otc_token' ('\<star\>') and its corresponding sentence-piece ('▁\<star\>') to 'words.txt' and 'tokens.txt,' respectively. Additionally, it computes SSL features using the 'wav2vec2-base' model. (You can use GPU to accelerate feature extraction).
+
+### Making synthetic errors to the transcript (train-clean-100) [optional]
+```
+sub_er=0.17
+ins_er=0.17
+del_er=0.17
+synthetic_train_manifest="librispeech_cuts_train-clean-100_${sub_er}_${ins_er}_${del_er}.jsonl.gz"
+
+./local/make_error_cutset.py \
+  --input-cutset "${manifest_dir}/librispeech_cuts_train-clean-100.jsonl.gz" \
+  --words-file "${lang_dir}/words.txt" \
+  --sub-error-rate "${sub_er}" \
+  --ins-error-rate "${ins_er}" \
+  --del-error-rate "${del_er}" \
+  --output-cutset "${manifest_dir}/${synthetic_train_manifest}"
+```
+This script generates synthetic substitution, insertion, and deletion errors in the transcript with ratios 'sub_er', 'ins_er', and 'del_er', respectively. The original transcript is saved as 'verbatim transcript' in the cutset, along with information on how the transcript is corrupted:
+
+  - '[hello]' indicates the original word 'hello' is substituted by another word
+  - '[]' indicates an extra word is inserted into the transcript
+  - '-hello-' indicates the word 'hello' is deleted from the transcript
+  
+So if the original transcript is "have a nice day" and the synthetic one is "a very good day", the 'verbatim transcript' would be:
+```
+original:  have  a      nice  day
+synthetic:       a very good  day
+verbatim: -have- a  [] [nice] day
+```
+
+### Training
+The training uses synthetic data based on the train-clean-100 subset.
+```
+otc_lang_dir=data/lang_bpe_200
+
+allow_bypass_arc=true
+allow_self_loop_arc=true
+initial_bypass_weight=-19
+initial_self_loop_weight=3.75
+bypass_weight_decay=0.975
+self_loop_weight_decay=0.999
+
+show_alignment=true
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./conformer_ctc2/train.py \
+  --world-size 4 \
+  --manifest-dir "${manifest_dir}" \
+  --train-manifest "${synthetic_train_manifest}" \
+  --exp-dir "${exp_dir}" \
+  --lang-dir "${otc_lang_dir}" \
+  --otc-token "${otc_token}" \
+  --allow-bypass-arc "${allow_bypass_arc}" \
+  --allow-self-loop-arc "${allow_self_loop_arc}" \
+  --initial-bypass-weight "${initial_bypass_weight}" \
+  --initial-self-loop-weight "${initial_self_loop_weight}" \
+  --bypass-weight-decay "${bypass_weight_decay}" \
+  --self-loop-weight-decay "${self_loop_weight_decay}" \
+  --show-alignment "${show_alignment}"
+```
+The bypass arc deals with substitution and insertion errors, while the self-loop arc deals with deletion errors. Using "--show-alignment" would print the best alignment during training, which is very helpful for tuning hyperparameters and debugging.
+
+### Decoding
+```
+export CUDA_VISIBLE_DEVICES="0"
+./conformer_ctc2/decode.py \
+  --manifest-dir "${manifest_dir}" \
+  --exp-dir "${exp_dir}" \
+  --lang-dir "${otc_lang_dir}" \
+  --lm-dir "${lm_dir}" \
+  --otc-token "${otc_token}"
+```
+
+### Results (ctc-greedy-search)
+<table>
+  <tr>
+    <td rowspan=2>Training Criterion</td>
+    <td colspan=2>ssl</td>
+    <td colspan=2>fbank</td>
+  </tr>
+  <tr>
+    <td>test-clean</td>
+    <td>test-other</td>
+    <td>test-clean</td>
+    <td>test-other</td>
+  </tr>
+  <tr>
+    <td>CTC</td>
+    <td>100.0</td>
+    <td>100.0</td>
+    <td>99.89</td>
+    <td>99.98</td>
+  </tr>
+  <tr>
+    <td>OTC</td>
+    <td>11.89</td>
+    <td>25.46</td>
+    <td>20.14</td>
+    <td>44.24</td>
+  </tr>
+</table>
+
+### Results (1best, blank_bias=-4)
+<table>
+  <tr>
+    <td rowspan=2>Training Criterion</td>
+    <td colspan=2>ssl</td>
+    <td colspan=2>fbank</td>
+  </tr>
+  <tr>
+    <td>test-clean</td>
+    <td>test-other</td>
+    <td>test-clean</td>
+    <td>test-other</td>
+  </tr>
+  <tr>
+    <td>CTC</td>
+    <td>98.40</td>
+    <td>98.68</td>
+    <td>99.79</td>
+    <td>99.86</td>
+  </tr>
+  <tr>
+    <td>OTC</td>
+    <td>6.59</td>
+    <td>15.98</td>
+    <td>11.78</td>
+    <td>32.38</td>
+  </tr>
+</table>
+
+## Pre-trained Model
+Pre-trained model: <https://huggingface.co/dgao/icefall-otc-librispeech-conformer-ctc2>
+
+## Citations
+```
+@inproceedings{gao2023bypass,
+  title={Bypass Temporal Classification: Weakly Supervised Automatic Speech Recognition with Imperfect Transcripts},
+  author={Gao, Dongji and Wiesner, Matthew and Xu, Hainan and Garcia, Leibny Paola and Povey, Daniel and Khudanpur, Sanjeev},
+  booktitle={INTERSPEECH},
+  year={2023}
+}
+
+@inproceedings{gao2023learning,
+  title={Learning from Flawed Data: Weakly Supervised Automatic Speech Recognition},
+  author={Gao, Dongji and Xu, Hainan and Raj, Desh and Garcia, Leibny Paola and Povey, Daniel and Khudanpur, Sanjeev},
+  booktitle={IEEE ASRU},
+  year={2023}
+}
+```
diff --git a/egs/librispeech/WSASR/conformer_ctc2/__init__.py b/egs/librispeech/WSASR/conformer_ctc2/__init__.py
new file mode 120000
index 000000000..43a85af20
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/__init__.py
@@ -0,0 +1 @@
+../../ASR/pruned_transducer_stateless2/__init__.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/conformer_ctc2/asr_datamodule.py b/egs/librispeech/WSASR/conformer_ctc2/asr_datamodule.py
new file mode 100644
index 000000000..2f8e658c5
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/asr_datamodule.py
@@ -0,0 +1,369 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+#                2023  John Hopkins University  (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import AudioSamples  # noqa F401 For AudioSamples
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class LibriSpeechAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--full-libri",
+            type=str2bool,
+            default=False,
+            help="""Used only when --mini-libri is False.When enabled,
+            use 960h LibriSpeech. Otherwise, use 100h subset.""",
+        )
+        group.add_argument(
+            "--mini-libri",
+            type=str2bool,
+            default=False,
+            help="True for mini librispeech",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/ssl"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+        group.add_argument(
+            "--train-manifest",
+            type=str,
+            default="librispeech_cuts_train-clean-100.jsonl.gz",
+            help="Train manifest file.",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+
+        validate = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_clean_5_cuts(self) -> CutSet:
+        logging.info("mini_librispeech: About to get train-clean-5 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-clean-5.jsonl.gz"
+        )
+
+    @lru_cache()
+    def train_clean_100_cuts(self) -> CutSet:
+        logging.info("About to get train-clean-100 cuts")
+        return load_manifest_lazy(self.args.manifest_dir / self.args.train_manifest)
+
+    @lru_cache()
+    def train_all_shuf_cuts(self) -> CutSet:
+        logging.info(
+            "About to get the shuffled train-clean-100, \
+            train-clean-360 and train-other-500 cuts"
+        )
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_train-all-shuf.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_clean_2_cuts(self) -> CutSet:
+        logging.info("mini_librispeech: About to get dev-clean-2 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-clean-2.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_clean_cuts(self) -> CutSet:
+        logging.info("About to get dev-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_other_cuts(self) -> CutSet:
+        logging.info("About to get dev-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_dev-other.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_clean_cuts(self) -> CutSet:
+        logging.info("About to get test-clean cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-clean.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_other_cuts(self) -> CutSet:
+        logging.info("About to get test-other cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "librispeech_cuts_test-other.jsonl.gz"
+        )
diff --git a/egs/librispeech/WSASR/conformer_ctc2/attention.py b/egs/librispeech/WSASR/conformer_ctc2/attention.py
new file mode 120000
index 000000000..e808a6f20
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/attention.py
@@ -0,0 +1 @@
+../../ASR/conformer_ctc2/attention.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/conformer_ctc2/conformer.py b/egs/librispeech/WSASR/conformer_ctc2/conformer.py
new file mode 100644
index 000000000..db4821d37
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/conformer.py
@@ -0,0 +1,949 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#                2022  Xiaomi Corp.                              (author: Quandong Wang)
+#                2023  Johns Hopkins University                  (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+import warnings
+from typing import Optional, Tuple
+
+import torch
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv1d,
+    ScaledLinear,
+)
+from subsampling import Conv2dSubsampling, Conv2dSubsampling2
+from torch import Tensor, nn
+from transformer import Supervisions, Transformer, encoder_padding_mask
+
+
+class Conformer(Transformer):
+    """
+    Args:
+        num_features (int): Number of input features
+        num_classes (int): Number of output classes
+        subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers)
+        d_model (int): attention dimension, also the output dimension
+        nhead (int): number of head
+        dim_feedforward (int): feedforward dimention
+        num_encoder_layers (int): number of encoder layers
+        num_decoder_layers (int): number of decoder layers
+        dropout (float): dropout rate
+        layer_dropout (float): layer-dropout rate.
+        cnn_module_kernel (int): Kernel size of convolution module
+        vgg_frontend (bool): whether to use vgg frontend.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 2,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.2,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+    ) -> None:
+        super(Conformer, self).__init__(
+            num_features=num_features,
+            num_classes=num_classes,
+            subsampling_factor=subsampling_factor,
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            num_encoder_layers=num_encoder_layers,
+            num_decoder_layers=num_decoder_layers,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+        )
+
+        self.num_features = num_features
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4 and subsampling_factor != 2:
+            raise NotImplementedError("Support only 'subsampling_factor=4 or 2'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        if self.subsampling_factor == 4:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+        elif self.subsampling_factor == 2:
+            self.encoder_embed = Conv2dSubsampling2(num_features, d_model)
+
+        self.encoder_pos = RelPositionalEncoding(d_model, dropout)
+
+        encoder_layer = ConformerEncoderLayer(
+            d_model,
+            nhead,
+            dim_feedforward,
+            dropout,
+            layer_dropout,
+            cnn_module_kernel,
+        )
+        self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        supervisions: Optional[Supervisions] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute encoder padding mask, which is used as memory key padding
+            mask for the decoder.
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+        Returns:
+            Tensor: Predictor tensor of dimension (input_length, batch_size, d_model).
+            Tensor: Mask tensor of dimension (batch_size, input_length)
+        """
+        x = self.encoder_embed(x)
+        x, pos_emb = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), self.subsampling_factor, supervisions)
+        if mask is not None:
+            mask = mask.to(x.device)
+
+        # Caution: We assume the subsampling factor is 4!
+
+        x = self.encoder(
+            x, pos_emb, src_key_padding_mask=mask, warmup=warmup
+        )  # (T, N, C)
+
+        # x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        # return x, lengths
+        return x, mask
+
+
+class ConformerEncoderLayer(nn.Module):
+    """
+    ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks.
+    See: "Conformer: Convolution-augmented Transformer for Speech Recognition"
+
+    Args:
+        d_model: the number of expected features in the input (required).
+        nhead: the number of heads in the multiheadattention models (required).
+        dim_feedforward: the dimension of the feedforward network model (default=2048).
+        dropout: the dropout value (default=0.1).
+        cnn_module_kernel (int): Kernel size of convolution module.
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = encoder_layer(src, pos_emb)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        cnn_module_kernel: int = 31,
+    ) -> None:
+        super(ConformerEncoderLayer, self).__init__()
+
+        self.layer_dropout = layer_dropout
+
+        self.d_model = d_model
+
+        self.self_attn = RelPositionMultiheadAttention(d_model, nhead, dropout=0.0)
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.feed_forward_macaron = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.conv_module = ConvolutionModule(d_model, cnn_module_kernel)
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            pos_emb: Positional embedding tensor (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # macaron style feed forward module
+        src = src + self.dropout(self.feed_forward_macaron(src))
+
+        # multi-headed self-attention module
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            pos_emb=pos_emb,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = src + self.dropout(src_att)
+
+        # convolution module
+        src = src + self.dropout(
+            self.conv_module(src, src_key_padding_mask=src_key_padding_mask)
+        )
+
+        # feed forward module
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+
+class ConformerEncoder(nn.Module):
+    r"""ConformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the ConformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8)
+        >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> pos_emb = torch.rand(32, 19, 512)
+        >>> out = conformer_encoder(src, pos_emb)
+    """
+
+    def __init__(self, encoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        src: Tensor,
+        pos_emb: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        warmup: float = 1.0,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            pos_emb: Positional embedding tensor (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            pos_emb: (N, 2*S-1, E)
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for i, mod in enumerate(self.layers):
+            output = mod(
+                output,
+                pos_emb,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+
+        return output
+
+
+class RelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module.
+
+    See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None:
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: Tensor) -> None:
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                # Note: TorchScript doesn't implement operator== for torch.Device
+                if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device):
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vecotr and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> Tuple[Tensor, Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Encoded tensor (batch, 2*time-1, `*`).
+
+        """
+        self.extend_pe(x)
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2
+            - x.size(1)
+            + 1 : self.pe.size(1) // 2  # noqa E203
+            + x.size(1),
+        ]
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class RelPositionMultiheadAttention(nn.Module):
+    r"""Multi-Head Attention layer with relative position encoding
+
+    See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context"
+
+    Args:
+        embed_dim: total dimension of the model.
+        num_heads: parallel attention heads.
+        dropout: a Dropout layer on attn_output_weights. Default: 0.0.
+
+    Examples::
+
+        >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb)
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+    ) -> None:
+        super(RelPositionMultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        self.in_proj = ScaledLinear(embed_dim, 3 * embed_dim, bias=True)
+        self.out_proj = ScaledLinear(
+            embed_dim, embed_dim, bias=True, initial_scale=0.25
+        )
+
+        # linear transformation for positional encoding.
+        self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim))
+        self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach())
+        self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach())
+        self._reset_parameters()
+
+    def _pos_bias_u(self):
+        return self.pos_bias_u * self.pos_bias_u_scale.exp()
+
+    def _pos_bias_v(self):
+        return self.pos_bias_v * self.pos_bias_v_scale.exp()
+
+    def _reset_parameters(self) -> None:
+        nn.init.normal_(self.pos_bias_u, std=0.01)
+        nn.init.normal_(self.pos_bias_v, std=0.01)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. When given a binary mask and a value is True,
+                the corresponding value on the attention layer will be ignored. When given
+                a byte mask and a value is non-zero, the corresponding value on the attention
+                layer will be ignored
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            - Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the position
+            with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            - Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+        return self.multi_head_attention_forward(
+            query,
+            key,
+            value,
+            pos_emb,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj.get_weight(),
+            self.in_proj.get_bias(),
+            self.dropout,
+            self.out_proj.get_weight(),
+            self.out_proj.get_bias(),
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+        )
+
+    def rel_shift(self, x: Tensor) -> Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x: Input tensor (batch, head, time1, 2*time1-1).
+                time1 means the length of query vector.
+
+        Returns:
+            Tensor: tensor of shape (batch, head, time1, time2)
+          (note: time2 has the same value as time1, but it is for
+          the key, while time1 is for the query).
+        """
+        (batch_size, num_heads, time1, n) = x.shape
+        assert n == 2 * time1 - 1
+        # Note: TorchScript requires explicit arg for stride()
+        batch_stride = x.stride(0)
+        head_stride = x.stride(1)
+        time1_stride = x.stride(2)
+        n_stride = x.stride(3)
+        return x.as_strided(
+            (batch_size, num_heads, time1, time1),
+            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+            storage_offset=n_stride * (time1 - 1),
+        )
+
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        pos_emb: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Tensor,
+        in_proj_bias: Tensor,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Tensor,
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query, key, value: map a query and a set of key-value pairs to an output.
+            pos_emb: Positional embedding tensor
+            embed_dim_to_check: total dimension of the model.
+            num_heads: parallel attention heads.
+            in_proj_weight, in_proj_bias: input projection weight and bias.
+            dropout_p: probability of an element to be zeroed.
+            out_proj_weight, out_proj_bias: the output projection weight and bias.
+            training: apply dropout if is ``True``.
+            key_padding_mask: if provided, specified padding elements in the key will
+                be ignored by the attention. This is an binary mask. When the value is True,
+                the corresponding value on the attention layer will be filled with -inf.
+            need_weights: output attn_output_weights.
+            attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+                the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+
+        Shape:
+            Inputs:
+            - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+            the embedding dimension.
+            - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+            the embedding dimension.
+            - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence
+            length, N is the batch size, E is the embedding dimension.
+            - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
+            If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
+            will be unchanged. If a BoolTensor is provided, the positions with the
+            value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+            - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+            3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+            S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+            positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
+            while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
+            are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+            is provided, it will be added to the attention weight.
+
+            Outputs:
+            - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+            E is the embedding dimension.
+            - attn_output_weights: :math:`(N, L, S)` where N is the batch size,
+            L is the target sequence length, S is the source sequence length.
+        """
+
+        tgt_len, bsz, embed_dim = query.size()
+        assert embed_dim == embed_dim_to_check
+        assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
+
+        head_dim = embed_dim // num_heads
+        assert (
+            head_dim * num_heads == embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        scaling = float(head_dim) ** -0.5
+
+        if torch.equal(query, key) and torch.equal(key, value):
+            # self-attention
+            q, k, v = nn.functional.linear(query, in_proj_weight, in_proj_bias).chunk(
+                3, dim=-1
+            )
+
+        elif torch.equal(key, value):
+            # encoder-decoder attention
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1)
+
+        else:
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = 0
+            _end = embed_dim
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            q = nn.functional.linear(query, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim
+            _end = embed_dim * 2
+            _w = in_proj_weight[_start:_end, :]
+            if _b is not None:
+                _b = _b[_start:_end]
+            k = nn.functional.linear(key, _w, _b)
+
+            # This is inline in_proj function with in_proj_weight and in_proj_bias
+            _b = in_proj_bias
+            _start = embed_dim * 2
+            _end = None
+            _w = in_proj_weight[_start:, :]
+            if _b is not None:
+                _b = _b[_start:]
+            v = nn.functional.linear(value, _w, _b)
+
+        if attn_mask is not None:
+            assert (
+                attn_mask.dtype == torch.float32
+                or attn_mask.dtype == torch.float64
+                or attn_mask.dtype == torch.float16
+                or attn_mask.dtype == torch.uint8
+                or attn_mask.dtype == torch.bool
+            ), "Only float, byte, and bool types are supported for attn_mask, not {}".format(
+                attn_mask.dtype
+            )
+            if attn_mask.dtype == torch.uint8:
+                warnings.warn(
+                    "Byte tensor for attn_mask is deprecated. Use bool tensor instead."
+                )
+                attn_mask = attn_mask.to(torch.bool)
+
+            if attn_mask.dim() == 2:
+                attn_mask = attn_mask.unsqueeze(0)
+                if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
+                    raise RuntimeError("The size of the 2D attn_mask is not correct.")
+            elif attn_mask.dim() == 3:
+                if list(attn_mask.size()) != [
+                    bsz * num_heads,
+                    query.size(0),
+                    key.size(0),
+                ]:
+                    raise RuntimeError("The size of the 3D attn_mask is not correct.")
+            else:
+                raise RuntimeError(
+                    "attn_mask's dimension {} is not supported".format(attn_mask.dim())
+                )
+            # attn_mask's dim is 3 now.
+
+        # convert ByteTensor key_padding_mask to bool
+        if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
+            warnings.warn(
+                "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead."
+            )
+            key_padding_mask = key_padding_mask.to(torch.bool)
+
+        q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim)
+        k = k.contiguous().view(-1, bsz, num_heads, head_dim)
+        v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
+
+        src_len = k.size(0)
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.size(0) == bsz, "{} == {}".format(
+                key_padding_mask.size(0), bsz
+            )
+            assert key_padding_mask.size(1) == src_len, "{} == {}".format(
+                key_padding_mask.size(1), src_len
+            )
+
+        q = q.transpose(0, 1)  # (batch, time1, head, d_k)
+
+        pos_emb_bsz = pos_emb.size(0)
+        assert pos_emb_bsz in (1, bsz)  # actually it is 1
+        p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        q_with_bias_u = (q + self._pos_bias_u()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        q_with_bias_v = (q + self._pos_bias_v()).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3
+        k = k.permute(1, 2, 3, 0)  # (batch, head, d_k, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k)  # (batch, head, time1, time2)
+
+        # compute matrix b and matrix d
+        matrix_bd = torch.matmul(
+            q_with_bias_v, p.transpose(-2, -1)
+        )  # (batch, head, time1, 2*time1-1)
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        attn_output_weights = matrix_ac + matrix_bd  # (batch, head, time1, time2)
+
+        attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, -1)
+
+        assert list(attn_output_weights.size()) == [
+            bsz * num_heads,
+            tgt_len,
+            src_len,
+        ]
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_output_weights.masked_fill_(attn_mask, float("-inf"))
+            else:
+                attn_output_weights += attn_mask
+
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            attn_output_weights = attn_output_weights.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2),
+                float("-inf"),
+            )
+            attn_output_weights = attn_output_weights.view(
+                bsz * num_heads, tgt_len, src_len
+            )
+
+        attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1)
+        attn_output_weights = nn.functional.dropout(
+            attn_output_weights, p=dropout_p, training=training
+        )
+
+        attn_output = torch.bmm(attn_output_weights, v)
+        assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
+        )
+        attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.view(
+                bsz, num_heads, tgt_len, src_len
+            )
+            return attn_output, attn_output_weights.sum(dim=1) / num_heads
+        else:
+            return attn_output, None
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py
+
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+        bias (bool): Whether to use bias in conv layers (default=True).
+
+    """
+
+    def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None:
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = ScaledConv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+
+        # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu).
+        # For most layers the normal rms value of channels of x seems to be in the range 1 to 4,
+        # but sometimes, for some reason, for layer 0 the rms ends up being very large,
+        # between 50 and 100 for different channels.  This will cause very peaky and
+        # sparse derivatives for the sigmoid gating function, which will tend to make
+        # the loss function not learn effectively.  (for most layers the average absolute values
+        # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion,
+        # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different
+        # layers, which likely breaks down as 0.5 for the "linear" half and
+        # 0.2 to 0.3 for the part that goes into the sigmoid.  The idea is that if we
+        # constrain the rms values to a reasonable range via a constraint of max_abs=10.0,
+        # it will be in a better position to start learning something, i.e. to latch onto
+        # the correct range.
+        self.deriv_balancer1 = ActivationBalancer(
+            channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0
+        )
+
+        self.depthwise_conv = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+
+        self.deriv_balancer2 = ActivationBalancer(
+            channel_dim=1, min_positive=0.05, max_positive=1.0
+        )
+
+        self.activation = DoubleSwish()
+
+        self.pointwise_conv2 = ScaledConv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+            initial_scale=0.25,
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Compute convolution module.
+
+        Args:
+            x: Input tensor (#time, batch, channels).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Returns:
+            Tensor: Output tensor (#time, batch, channels).
+
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.permute(1, 2, 0)  # (#batch, channels, time).
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channels, time)
+
+        x = self.deriv_balancer1(x)
+        x = nn.functional.glu(x, dim=1)  # (batch, channels, time)
+
+        # 1D Depthwise Conv
+        if src_key_padding_mask is not None:
+            x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0)
+        x = self.depthwise_conv(x)
+
+        x = self.deriv_balancer2(x)
+        x = self.activation(x)
+
+        x = self.pointwise_conv2(x)  # (batch, channel, time)
+
+        return x.permute(2, 0, 1)
+
+
+if __name__ == "__main__":
+    feature_dim = 50
+    c = Conformer(num_features=feature_dim, d_model=128, nhead=4)
+    batch_size = 5
+    seq_len = 20
+    # Just make sure the forward pass runs.
+    f = c(
+        torch.randn(batch_size, seq_len, feature_dim),
+        torch.full((batch_size,), seq_len, dtype=torch.int64),
+        warmup=0.5,
+    )
diff --git a/egs/librispeech/WSASR/conformer_ctc2/decode.py b/egs/librispeech/WSASR/conformer_ctc2/decode.py
new file mode 100755
index 000000000..3fa045533
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/decode.py
@@ -0,0 +1,718 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo,
+#                                            Fangjun Kuang,
+#                                            Quandong Wang)
+#           2023 Johns Hopkins University (Author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.decode import get_lattice, one_best_decoding
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.otc_graph_compiler import OtcTrainingGraphCompiler
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    load_averaged_model,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--otc-token",
+        type=str,
+        default="<star>",
+        help="OTC token",
+    )
+
+    parser.add_argument(
+        "--blank-bias",
+        type=float,
+        default=0,
+        help="bias (log-prob) added to blank token during decoding",
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=1,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="ctc-greedy-search",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) ctc-greedy-search. It only use CTC output and a sentence piece
+              model for decoding. It produces the same results with ctc-decoding.
+            - (2) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+       """,
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=0,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc2/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_200",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "subsampling_factor": 2,
+            "feature_dim": 768,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "encoder_dim": 512,
+            "num_encoder_layers": 12,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def ctc_greedy_search(
+    nnet_output: torch.Tensor,
+    memory: torch.Tensor,
+    memory_key_padding_mask: torch.Tensor,
+) -> List[List[int]]:
+    """Apply CTC greedy search
+
+     Args:
+         speech (torch.Tensor): (batch, max_len, feat_dim)
+         speech_length (torch.Tensor): (batch, )
+    Returns:
+         List[List[int]]: best path result
+    """
+    batch_size = memory.shape[1]
+    # Let's assume B = batch_size
+    encoder_out = memory
+    encoder_mask = memory_key_padding_mask
+    maxlen = encoder_out.size(0)
+
+    ctc_probs = nnet_output  # (B, maxlen, vocab_size)
+    topk_prob, topk_index = ctc_probs.topk(1, dim=2)  # (B, maxlen, 1)
+    topk_index = topk_index.view(batch_size, maxlen)  # (B, maxlen)
+    topk_index = topk_index.masked_fill_(encoder_mask, 0)  # (B, maxlen)
+    hyps = [hyp.tolist() for hyp in topk_index]
+    scores = topk_prob.max(1)
+    hyps = [remove_duplicates_and_blank(hyp) for hyp in hyps]
+    return hyps, scores
+
+
+def remove_duplicates_and_blank(hyp: List[int]) -> List[int]:
+    # from https://github.com/wenet-e2e/wenet/blob/main/wenet/utils/common.py
+    new_hyp: List[int] = []
+    cur = 0
+    while cur < len(hyp):
+        if hyp[cur] != 0:
+            new_hyp.append(hyp[cur])
+        prev = cur
+        while cur < len(hyp) and hyp[cur] == hyp[prev]:
+            cur += 1
+    return new_hyp
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+    nnet_output[:, :, 0] += params.blank_bias
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                params.subsampling_factor,
+                rounding_mode="trunc",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                params.subsampling_factor,
+                rounding_mode="trunc",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor + 2,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.method == "ctc-greedy-search":
+        hyps, _ = ctc_greedy_search(
+            nnet_output,
+            memory,
+            memory_key_padding_mask,
+        )
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(hyps)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-greedy-search"
+        return {key: hyps}
+
+    if params.method in ["1best"]:
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        key = "no_rescore"
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+
+        return {key: hyps}
+    else:
+        assert False, f"Unsupported decoding method: {params.method}"
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, hyps in hyps_dict.items():
+                this_batch = []
+                assert len(hyps) == len(texts)
+                for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                    ref_words = ref_text.split()
+                    this_batch.append((cut_id, ref_words, hyp_words))
+
+                results[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for ref_text in texts:
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    if params.method in ("attention-decoder", "rnn-lm"):
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = params.exp_dir / f"errs-{test_set_name}-{key}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = params.exp_dir / f"wer-summary-{test_set_name}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+    assert "▁" not in args.otc_token
+    args.otc_token = f"▁{args.otc_token}"
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    # remove otc_token from decoding units
+    max_token_id = max(lexicon.tokens) - 1
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = OtcTrainingGraphCompiler(
+        params.lang_dir,
+        params.otc_token,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    params.num_classes = num_classes
+    params.sos_id = sos_id
+    params.eos_id = eos_id
+
+    if params.method == "ctc-decoding" or params.method == "ctc-greedy-search":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    G = None
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.encoder_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+    )
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/conformer_ctc2/export.py b/egs/librispeech/WSASR/conformer_ctc2/export.py
new file mode 120000
index 000000000..5f484e391
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/export.py
@@ -0,0 +1 @@
+../../ASR/conformer_ctc2/export.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/conformer_ctc2/label_smoothing.py b/egs/librispeech/WSASR/conformer_ctc2/label_smoothing.py
new file mode 120000
index 000000000..c050ea637
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/label_smoothing.py
@@ -0,0 +1 @@
+../../ASR/conformer_ctc/label_smoothing.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/conformer_ctc2/optim.py b/egs/librispeech/WSASR/conformer_ctc2/optim.py
new file mode 120000
index 000000000..db836b5e0
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/optim.py
@@ -0,0 +1 @@
+../../ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/conformer_ctc2/scaling.py b/egs/librispeech/WSASR/conformer_ctc2/scaling.py
new file mode 120000
index 000000000..bd0abfeee
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/scaling.py
@@ -0,0 +1 @@
+../../ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/librispeech/WSASR/conformer_ctc2/subsampling.py b/egs/librispeech/WSASR/conformer_ctc2/subsampling.py
new file mode 100644
index 000000000..2ba802866
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/subsampling.py
@@ -0,0 +1,184 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  University of Chinese Academy of Sciences (author: Han Zhu)
+#                2022  Xiaomi Corporation                        (author: Quandong Wang)
+#                2023  Johns Hopkins University                  (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledConv2d,
+    ScaledLinear,
+)
+
+
+class Conv2dSubsampling(torch.nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        """
+        Args:
+          in_channels:
+            Number of channels in. The input shape is (N, T, in_channels).
+            Caution: It requires: T >=7, in_channels >=7
+          out_channels
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels)
+          layer1_channels:
+            Number of channels in layer1
+          layer1_channels:
+            Number of channels in layer2
+        """
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = torch.nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels
+        )
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
+
+
+class Conv2dSubsampling2(torch.nn.Module):
+    """Convolutional 2D subsampling (to 1/2 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim) where
+    T' = (T - 1) // 2 - 2, which approximates T' == T // 2
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        layer1_channels: int = 8,
+        layer2_channels: int = 32,
+        layer3_channels: int = 128,
+    ) -> None:
+        assert in_channels >= 7
+        super().__init__()
+
+        self.conv = torch.nn.Sequential(
+            ScaledConv2d(
+                in_channels=1,
+                out_channels=layer1_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer1_channels,
+                out_channels=layer2_channels,
+                kernel_size=3,
+                stride=2,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=layer2_channels,
+                out_channels=layer3_channels,
+                kernel_size=3,
+                stride=1,
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(
+            layer3_channels * ((in_channels - 1) // 2 - 2), out_channels
+        )
+        self.out_norm = BasicNorm(out_channels, learn_eps=False)
+        self.out_balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = x.unsqueeze(1)
+        x = self.conv(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
diff --git a/egs/librispeech/WSASR/conformer_ctc2/train.py b/egs/librispeech/WSASR/conformer_ctc2/train.py
new file mode 100755
index 000000000..fe6c5af91
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/train.py
@@ -0,0 +1,1115 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang,
+#                                                  Mingshuang Luo,
+#                                                  Zengwei Yao,
+#                                                  Quandong Wang)
+#              2023 Johns Hopkins University (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./conformer_ctc2/train.py \
+  --world-size 4 \
+  --manifest-dir data/ssl \
+  --train-manifest librispeech_cuts_train-clean-100_0.17_0.17_0.17.jsonl.gz \
+  --exp-dir conformer_ctc2/exp \
+  --lang-dir data/lang_bpe_200 \
+  --otc-token "<star>" \
+  --allow-bypass-arc true \
+  --allow-self-loop-arc true \
+  --initial-bypass-weight -19 \
+  --initial-self-loop-weight 3.75 \
+  --bypass-weight-decay 0.975 \
+  --self-loop-weight-decay 0.999 \
+  --show-alignment true
+"""
+
+
+import argparse
+import copy
+import logging
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.decode import one_best_decoding
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.otc_graph_compiler import OtcTrainingGraphCompiler
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions_otc,
+    get_texts,
+    setup_logger,
+    str2bool,
+)
+
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=20,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc2/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_200",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="""The initial learning rate. This value should not need to be
+        changed.""",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate decreases.
+        We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.0,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * ctc_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=0,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=8000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=10,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--otc-token",
+        type=str,
+        default="_<star>",
+        help="OTC token",
+    )
+
+    parser.add_argument(
+        "--allow-bypass-arc",
+        type=str2bool,
+        default=True,
+        help="""Whether to add bypass arc to training graph for substitution
+        and insertion errors (wrong or extra words in the transcript).""",
+    )
+
+    parser.add_argument(
+        "--allow-self-loop-arc",
+        type=str2bool,
+        default=True,
+        help="""Whether to self-loop bypass arc to training graph for deletion errors
+        (missing words in the transcript).""",
+    )
+
+    parser.add_argument(
+        "--initial-bypass-weight",
+        type=float,
+        default=0.0,
+        help="Initial weight associated with bypass arc",
+    )
+
+    parser.add_argument(
+        "--initial-self-loop-weight",
+        type=float,
+        default=0.0,
+        help="Initial weight associated with self-loop arc",
+    )
+
+    parser.add_argument(
+        "--bypass-weight-decay",
+        type=float,
+        default=1.0,
+        help="""Weight decay factor of bypass arc weight:
+        bypass_arc_weight = intial_bypass_weight * bypass_weight_decay ^ ith-epoch""",
+    )
+
+    parser.add_argument(
+        "--self-loop-weight-decay",
+        type=float,
+        default=1.0,
+        help="""Weight decay factor of self-loop arc weight:
+        self_loop_arc_weight = intial_self_loop_weight * self_loop_weight_decay ^ ith-epoch""",
+    )
+
+    parser.add_argument(
+        "--show-alignment",
+        type=str2bool,
+        default=True,
+        help="Whether to print OTC alignment during training",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 1,
+            "reset_interval": 200,
+            "valid_interval": 800,  # For the 100h subset, use 800
+            "alignment_interval": 25,
+            # parameters for conformer
+            "feature_dim": 768,
+            "subsampling_factor": 2,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for ctc loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    batch: dict,
+    graph_compiler: OtcTrainingGraphCompiler,
+    is_training: bool,
+    warmup: float = 2.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute OTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(
+            feature, supervisions, warmup=warmup
+        )
+        # Set the probability of OTC token as the average of non-blank tokens
+        # under the assumption that blank is the first and
+        # OTC token is the last token in tokens.txt
+        _, _, V = nnet_output.shape
+
+        otc_token_log_prob = torch.logsumexp(
+            nnet_output[:, :, 1:], dim=-1, keepdim=True
+        ) - torch.log(torch.tensor([V - 1])).to(device)
+
+        nnet_output = torch.cat([nnet_output, otc_token_log_prob], dim=-1)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts, utt_ids, verbatim_texts = encode_supervisions_otc(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+
+    bypass_weight = graph_compiler.initial_bypass_weight * (
+        graph_compiler.bypass_weight_decay ** (params.cur_epoch - 1)
+    )
+    self_loop_weight = graph_compiler.initial_self_loop_weight * (
+        graph_compiler.self_loop_weight_decay ** (params.cur_epoch - 1)
+    )
+
+    decoding_graph = graph_compiler.compile(
+        texts=texts,
+        allow_bypass_arc=params.allow_bypass_arc,
+        allow_self_loop_arc=params.allow_self_loop_arc,
+        bypass_weight=bypass_weight,
+        self_loop_weight=self_loop_weight,
+    )
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=3,
+    )
+
+    otc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    assert params.att_rate == 0.0
+    loss = otc_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+    info["otc_loss"] = otc_loss.detach().cpu().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = feature_lens.sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
+    )
+
+    if params.show_alignment:
+        if params.batch_idx_train % params.alignment_interval == 0:
+            for index, utt_id in enumerate(utt_ids):
+                verbatim_text = verbatim_texts[index]
+                utt_id = utt_ids[index]
+
+                lattice = k2.intersect_dense(
+                    decoding_graph,
+                    dense_fsa_vec,
+                    params.beam_size,
+                )
+                best_path = one_best_decoding(
+                    lattice=lattice,
+                    use_double_scores=params.use_double_scores,
+                )
+                hyp_ids = get_texts(best_path)[index]
+                hyp_text_list = [graph_compiler.token_table[i] for i in hyp_ids]
+                hyp_text = "".join(hyp_text_list).replace("▁", " ")
+
+                logging.info(f"[utterance id]: {utt_id}")
+                logging.info(f"[verbatim text]: {verbatim_text}")
+                logging.info(f"[best alignment]: {hyp_text}")
+                logging.info(bypass_weight)
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: OtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: OtcTrainingGraphCompiler,
+    scheduler: LRSchedulerType,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                params=params,
+                model=model,
+                batch=batch,
+                graph_compiler=graph_compiler,
+                is_training=True,
+                warmup=(params.batch_idx_train / params.model_warm_step),
+            )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        # scaler.scale(loss).backward()
+
+        try:
+            # loss.backward()
+            scaler.scale(loss).backward()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(f"failing batch size:{batch_size} ")
+            raise
+
+        scheduler.step_batch(params.batch_idx_train)
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+        if params.print_diagnostics and batch_idx == 30:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+            if loss_info["otc_loss"] == float("inf"):
+                logging.error("Your loss contains inf, something goes wrong")
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    params.valid_interval = 1600
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    graph_compiler = OtcTrainingGraphCompiler(
+        params.lang_dir,
+        otc_token=params.otc_token,
+        device=device,
+        initial_bypass_weight=params.initial_bypass_weight,
+        initial_self_loop_weight=params.initial_self_loop_weight,
+        bypass_weight_decay=params.bypass_weight_decay,
+        self_loop_weight_decay=params.self_loop_weight_decay,
+    )
+
+    # remove OTC token as it is the average of all non-blank tokens
+    max_token_id = graph_compiler.get_max_token_id() - 1
+    # add blank
+    num_classes = max_token_id + 1
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.encoder_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_encoder_layers=params.num_encoder_layers,
+        num_decoder_layers=params.num_decoder_layers,
+    )
+
+    print(model)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 20.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            scheduler=scheduler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: OtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    batch=batch,
+                    graph_compiler=graph_compiler,
+                    is_training=True,
+                    warmup=0.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    assert "▁" not in args.otc_token
+    args.otc_token = f"▁{args.otc_token}"
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/conformer_ctc2/transformer.py b/egs/librispeech/WSASR/conformer_ctc2/transformer.py
new file mode 100644
index 000000000..41e6cd357
--- /dev/null
+++ b/egs/librispeech/WSASR/conformer_ctc2/transformer.py
@@ -0,0 +1,1055 @@
+# Copyright    2021  University of Chinese Academy of Sciences (author: Han Zhu)
+# Copyright    2022  Xiaomi Corp.                              (author: Quandong Wang)
+#              2023  Johns Hopkins University                  (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import math
+from typing import Dict, List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from attention import MultiheadAttention
+from label_smoothing import LabelSmoothingLoss
+from scaling import (
+    ActivationBalancer,
+    BasicNorm,
+    DoubleSwish,
+    ScaledEmbedding,
+    ScaledLinear,
+)
+from subsampling import Conv2dSubsampling
+from torch.nn.utils.rnn import pad_sequence
+
+# Note: TorchScript requires Dict/List/etc. to be fully typed.
+Supervisions = Dict[str, torch.Tensor]
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        num_features: int,
+        num_classes: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        num_decoder_layers: int = 6,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          num_classes:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder/decoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          num_decoder_layers:
+            Number of decoder layers.
+          dropout:
+            Dropout in encoder/decoder.
+          layer_dropout (float): layer-dropout rate.
+        """
+        super().__init__()
+
+        self.num_features = num_features
+        self.num_classes = num_classes
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4 and subsampling_factor != 2:
+            raise NotImplementedError("Support only 'subsampling_factor=4 or 2'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_classes)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_classes -> d_model
+        self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            layer_dropout=layer_dropout,
+        )
+
+        self.encoder = TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), ScaledLinear(d_model, num_classes, bias=True)
+        )
+
+        if num_decoder_layers > 0:
+            self.decoder_num_class = (
+                self.num_classes
+            )  # bpe model already has sos/eos symbol
+
+            self.decoder_embed = ScaledEmbedding(
+                num_embeddings=self.decoder_num_class, embedding_dim=d_model
+            )
+            self.decoder_pos = PositionalEncoding(d_model, dropout)
+
+            decoder_layer = TransformerDecoderLayer(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+            )
+
+            self.decoder = TransformerDecoder(
+                decoder_layer=decoder_layer,
+                num_layers=num_decoder_layers,
+            )
+
+            self.decoder_output_layer = ScaledLinear(
+                d_model, self.decoder_num_class, bias=True
+            )
+
+            self.decoder_criterion = LabelSmoothingLoss()
+        else:
+            self.decoder_criterion = None
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        supervision: Optional[Supervisions] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (N, T, C).
+          supervision:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            (CAUTION: It contains length information, i.e., start and number of
+             frames, before subsampling)
+          warmup:
+            A floating point value that gradually increases from 0 throughout
+            training; when it is >= 1.0 we are "fully warmed up".  It is used
+            to turn modules on sequentially.
+
+        Returns:
+          Return a tuple containing 3 tensors:
+            - CTC output for ctc decoding. Its shape is (N, T, C)
+            - Encoder output with shape (T, N, C). It can be used as key and
+              value for the decoder.
+            - Encoder output padding mask. It can be used as
+              memory_key_padding_mask for the decoder. Its shape is (N, T).
+              It is None if `supervision` is None.
+        """
+
+        encoder_memory, memory_key_padding_mask = self.run_encoder(
+            x, supervision, warmup
+        )
+
+        x = self.ctc_output(encoder_memory)
+        return x, encoder_memory, memory_key_padding_mask
+
+    def run_encoder(
+        self,
+        x: torch.Tensor,
+        supervisions: Optional[Supervisions] = None,
+        warmup: float = 1.0,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Run the transformer encoder.
+
+        Args:
+          x:
+            The model input. Its shape is (N, T, C).
+          supervisions:
+            Supervision in lhotse format.
+            See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+            CAUTION: It contains length information, i.e., start and number of
+            frames, before subsampling
+            It is read directly from the batch, without any sorting. It is used
+            to compute the encoder padding mask, which is used as memory key
+            padding mask for the decoder.
+        Returns:
+          Return a tuple with two tensors:
+            - The encoder output, with shape (T, N, C)
+            - encoder padding mask, with shape (N, T).
+              The mask is None if `supervisions` is None.
+              It is used as memory key padding mask in the decoder.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        mask = encoder_padding_mask(x.size(0), supervisions)
+        mask = mask.to(x.device) if mask is not None else None
+        x = self.encoder(x, src_key_padding_mask=mask, warmup=warmup)  # (T, N, C)
+
+        return x, mask
+
+    def ctc_output(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          x:
+            The output tensor from the transformer encoder.
+            Its shape is (T, N, C)
+
+        Returns:
+          Return a tensor that can be used for CTC decoding.
+          Its shape is (N, T, C)
+        """
+        x = self.encoder_output_layer(x)
+        x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+        x = nn.functional.log_softmax(x, dim=-1)  # (N, T, C)
+        return x
+
+    @torch.jit.export
+    def decoder_forward(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[List[int]],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs. Each sublist contains IDs for an utterance.
+            The IDs can be either phone IDs or word piece IDs.
+          sos_id:
+            sos token id
+          eos_id:
+            eos token id
+
+        Returns:
+            A scalar, the **sum** of label smoothing loss over utterances
+            in the batch without any normalization.
+        """
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device)
+        ys_out_pad = ys_out_pad.to(device)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (N, T) -> (N, T, C)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, N, C)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, N, C) -> (N, T, C)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (N, T, C)
+
+        decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad)
+
+        return decoder_loss
+
+    @torch.jit.export
+    def decoder_nll(
+        self,
+        memory: torch.Tensor,
+        memory_key_padding_mask: torch.Tensor,
+        token_ids: List[torch.Tensor],
+        sos_id: int,
+        eos_id: int,
+    ) -> torch.Tensor:
+        """
+        Args:
+          memory:
+            It's the output of the encoder with shape (T, N, C)
+          memory_key_padding_mask:
+            The padding mask from the encoder.
+          token_ids:
+            A list-of-list IDs (e.g., word piece IDs).
+            Each sublist represents an utterance.
+          sos_id:
+            The token ID for SOS.
+          eos_id:
+            The token ID for EOS.
+        Returns:
+            A 2-D tensor of shape (len(token_ids), max_token_length)
+            representing the cross entropy loss (i.e., negative log-likelihood).
+        """
+        # The common part between this function and decoder_forward could be
+        # extracted as a separate function.
+        if isinstance(token_ids[0], torch.Tensor):
+            # This branch is executed by torchscript in C++.
+            # See https://github.com/k2-fsa/k2/pull/870
+            # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286
+            token_ids = [tolist(t) for t in token_ids]
+
+        ys_in = add_sos(token_ids, sos_id=sos_id)
+        ys_in = [torch.tensor(y) for y in ys_in]
+        ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id))
+
+        ys_out = add_eos(token_ids, eos_id=eos_id)
+        ys_out = [torch.tensor(y) for y in ys_out]
+        ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1))
+
+        device = memory.device
+        ys_in_pad = ys_in_pad.to(device, dtype=torch.int64)
+        ys_out_pad = ys_out_pad.to(device, dtype=torch.int64)
+
+        tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device)
+
+        tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id)
+        # TODO: Use length information to create the decoder padding mask
+        # We set the first column to False since the first column in ys_in_pad
+        # contains sos_id, which is the same as eos_id in our current setting.
+        tgt_key_padding_mask[:, 0] = False
+
+        tgt = self.decoder_embed(ys_in_pad)  # (B, T) -> (B, T, F)
+        tgt = self.decoder_pos(tgt)
+        tgt = tgt.permute(1, 0, 2)  # (B, T, F) -> (T, B, F)
+        pred_pad = self.decoder(
+            tgt=tgt,
+            memory=memory,
+            tgt_mask=tgt_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+        )  # (T, B, F)
+        pred_pad = pred_pad.permute(1, 0, 2)  # (T, B, F) -> (B, T, F)
+        pred_pad = self.decoder_output_layer(pred_pad)  # (B, T, F)
+        # nll: negative log-likelihood
+        nll = torch.nn.functional.cross_entropy(
+            pred_pad.view(-1, self.decoder_num_class),
+            ys_out_pad.view(-1),
+            ignore_index=-1,
+            reduction="none",
+        )
+
+        nll = nll.view(pred_pad.shape[0], -1)
+
+        return nll
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+
+        self.layer_dropout = layer_dropout
+
+        self.self_attn = MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+            warmup: controls selective bypass of of layers; if < 1.0, we will
+              bypass layers more frequently.
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        src_orig = src
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # src_att = self.self_attn(src, src, src, src_mask)
+        src_att = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = src + self.dropout(src_att)
+
+        src = src + self.dropout(self.feed_forward(src))
+
+        src = self.norm_final(self.balancer(src))
+
+        if alpha != 1.0:
+            src = alpha * src + (1 - alpha) * src_orig
+
+        return src
+
+
+class TransformerDecoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerDecoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+
+    Examples::
+        >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(20, 32, 512)
+        >>> out = decoder_layer(tgt, memory)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        layer_dropout: float = 0.075,
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerDecoderLayer, self).__init__()
+        self.layer_dropout = layer_dropout
+        self.self_attn = MultiheadAttention(d_model, nhead, dropout=0.0)
+        self.src_attn = MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.feed_forward = nn.Sequential(
+            ScaledLinear(d_model, dim_feedforward),
+            ActivationBalancer(channel_dim=-1),
+            DoubleSwish(),
+            nn.Dropout(dropout),
+            ScaledLinear(dim_feedforward, d_model, initial_scale=0.25),
+        )
+
+        self.norm_final = BasicNorm(d_model)
+
+        # try to ensure the output is close to zero-mean (or at least, zero-median).
+        self.balancer = ActivationBalancer(
+            channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0
+        )
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+          tgt:
+            the sequence to the decoder layer (required).
+          memory:
+            the sequence from the last layer of the encoder (required).
+          tgt_mask:
+            the mask for the tgt sequence (optional).
+          memory_mask:
+            the mask for the memory sequence (optional).
+          tgt_key_padding_mask:
+            the mask for the tgt keys per batch (optional).
+          memory_key_padding_mask:
+            the mask for the memory keys per batch (optional).
+          warmup: controls selective bypass of of layers; if < 1.0, we will
+            bypass layers more frequently.
+
+
+
+        Shape:
+            tgt: (T, N, E).
+            memory: (S, N, E).
+            tgt_mask: (T, T).
+            memory_mask: (T, S).
+            tgt_key_padding_mask: (N, T).
+            memory_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        tgt_orig = tgt
+
+        warmup_scale = min(0.1 + warmup, 1.0)
+        # alpha = 1.0 means fully use this encoder layer, 0.0 would mean
+        # completely bypass it.
+        if self.training:
+            alpha = (
+                warmup_scale
+                if torch.rand(()).item() <= (1.0 - self.layer_dropout)
+                else 0.1
+            )
+        else:
+            alpha = 1.0
+
+        # tgt_att = self.self_attn(tgt, tgt, tgt, tgt_mask)
+        tgt_att = self.self_attn(
+            tgt,
+            tgt,
+            tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout(tgt_att)
+
+        # src_att = self.src_attn(tgt, memory, memory, memory_mask)
+        src_att = self.src_attn(
+            tgt,
+            memory,
+            memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout(src_att)
+
+        tgt = tgt + self.dropout(self.feed_forward(tgt))
+
+        tgt = self.norm_final(self.balancer(tgt))
+
+        if alpha != 1.0:
+            tgt = alpha * tgt + (1 - alpha) * tgt_orig
+
+        return tgt
+
+
+class TransformerEncoder(nn.Module):
+    r"""TransformerEncoder is a stack of N encoder layers
+
+    Args:
+        encoder_layer: an instance of the TransformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> transformer_encoder = TransformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = transformer_encoder(src)
+    """
+
+    def __init__(self, encoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(encoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            src: (S, N, E).
+            mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number
+
+        """
+        output = src
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                warmup=warmup,
+            )
+
+        return output
+
+
+class TransformerDecoder(nn.Module):
+    r"""TransformerDecoder is a stack of N decoder layers
+
+    Args:
+        decoder_layer: an instance of the TransformerDecoderLayer() class (required).
+        num_layers: the number of sub-decoder-layers in the decoder (required).
+
+    Examples::
+        >>> decoder_layer = TransformerDecoderLayer(d_model=512, nhead=8)
+        >>> transformer_decoder = TransformerDecoder(decoder_layer, num_layers=6)
+        >>> memory = torch.rand(10, 32, 512)
+        >>> tgt = torch.rand(10, 32, 512)
+        >>> out = transformer_decoder(tgt, memory)
+    """
+
+    def __init__(self, decoder_layer: nn.Module, num_layers: int) -> None:
+        super().__init__()
+        self.layers = nn.ModuleList(
+            [copy.deepcopy(decoder_layer) for i in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        memory: torch.Tensor,
+        tgt_mask: Optional[torch.Tensor] = None,
+        memory_mask: Optional[torch.Tensor] = None,
+        tgt_key_padding_mask: Optional[torch.Tensor] = None,
+        memory_key_padding_mask: Optional[torch.Tensor] = None,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        r"""Pass the input through the decoder layers in turn.
+
+        Args:
+            tgt: the sequence to the decoder (required).
+            memory: the sequence from the last layer of the encoder (required).
+            tgt_mask: the mask for the tgt sequence (optional).
+            memory_mask: the mask for the memory sequence (optional).
+            tgt_key_padding_mask: the mask for the tgt keys per batch (optional).
+            memory_key_padding_mask: the mask for the memory keys per batch (optional).
+
+        Shape:
+            tgt: (S, N, E).
+            tgt_mask: (S, S).
+            tgt_key_padding_mask: (N, S).
+
+        """
+        output = tgt
+
+        for mod in self.layers:
+            output = mod(
+                output,
+                memory,
+                tgt_mask=tgt_mask,
+                memory_mask=memory_mask,
+                tgt_key_padding_mask=tgt_key_padding_mask,
+                memory_key_padding_mask=memory_key_padding_mask,
+                warmup=warmup,
+            )
+
+        return output
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)
+
+
+def encoder_padding_mask(
+    max_len: int,
+    subsampling_factor: Optional[int] = 4,
+    supervisions: Optional[Supervisions] = None,
+) -> Optional[torch.Tensor]:
+    """Make mask tensor containing indexes of padded part.
+
+    TODO::
+      This function **assumes** that the model uses
+      a subsampling factor of 4 or 2. We should remove that
+      assumption later.
+
+    Args:
+      max_len:
+        Maximum length of input features.
+        CAUTION: It is the length after subsampling.
+      supervisions:
+        Supervision in lhotse format.
+        See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32  # noqa
+        (CAUTION: It contains length information, i.e., start and number of
+         frames, before subsampling)
+
+    Returns:
+        Tensor: Mask tensor of dimension (batch_size, input_length),
+        True denote the masked indices.
+    """
+    if supervisions is None:
+        return None
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"],
+            supervisions["num_frames"],
+        ),
+        1,
+    ).to(torch.int32)
+
+    lengths = [0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)]
+    for idx in range(supervision_segments.size(0)):
+        # Note: TorchScript doesn't allow to unpack tensors as tuples
+        sequence_idx = supervision_segments[idx, 0].item()
+        start_frame = supervision_segments[idx, 1].item()
+        num_frames = supervision_segments[idx, 2].item()
+        lengths[sequence_idx] = start_frame + num_frames
+
+    if subsampling_factor == 4:
+        lengths = [((i - 1) // 2 - 1) // 2 for i in lengths]
+    elif subsampling_factor == 2:
+        lengths = [(i - 1) // 2 - 2 for i in lengths]
+    bs = int(len(lengths))
+    seq_range = torch.arange(0, max_len, dtype=torch.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len)
+    # Note: TorchScript doesn't implement Tensor.new()
+    seq_length_expand = torch.tensor(
+        lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype
+    ).unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+
+    return mask
+
+
+def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor:
+    """Generate a length mask for input.
+
+    The masked position are filled with True,
+    Unmasked positions are filled with False.
+
+    Args:
+      ys_pad:
+        padded tensor of dimension (batch_size, input_length).
+      ignore_id:
+        the ignored number (the padding number) in ys_pad
+
+    Returns:
+      Tensor:
+        a bool tensor of the same shape as the input tensor.
+    """
+    ys_mask = ys_pad == ignore_id
+    return ys_mask
+
+
+def generate_square_subsequent_mask(sz: int) -> torch.Tensor:
+    """Generate a square mask for the sequence. The masked positions are
+    filled with float('-inf'). Unmasked positions are filled with float(0.0).
+    The mask can be used for masked self-attention.
+
+    For instance, if sz is 3, it returns::
+
+        tensor([[0., -inf, -inf],
+                [0., 0., -inf],
+                [0., 0., 0]])
+
+    Args:
+      sz: mask size
+
+    Returns:
+      A square mask of dimension (sz, sz)
+    """
+    mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+    mask = (
+        mask.float()
+        .masked_fill(mask == 0, float("-inf"))
+        .masked_fill(mask == 1, float(0.0))
+    )
+    return mask
+
+
+def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]:
+    """Prepend sos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      sos_id:
+        The ID of the SOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist starts
+      with SOS ID.
+    """
+    return [[sos_id] + utt for utt in token_ids]
+
+
+def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]:
+    """Append eos_id to each utterance.
+
+    Args:
+      token_ids:
+        A list-of-list of token IDs. Each sublist contains
+        token IDs (e.g., word piece IDs) of an utterance.
+      eos_id:
+        The ID of the EOS token.
+
+    Return:
+      Return a new list-of-list, where each sublist ends
+      with EOS ID.
+    """
+    return [utt + [eos_id] for utt in token_ids]
+
+
+def tolist(t: torch.Tensor) -> List[int]:
+    """Used by jit"""
+    return torch.jit.annotate(List[int], t.tolist())
diff --git a/egs/librispeech/WSASR/figures/del.png b/egs/librispeech/WSASR/figures/del.png
new file mode 100644
index 000000000..38973980b
Binary files /dev/null and b/egs/librispeech/WSASR/figures/del.png differ
diff --git a/egs/librispeech/WSASR/figures/ins.png b/egs/librispeech/WSASR/figures/ins.png
new file mode 100644
index 000000000..2d0e807a9
Binary files /dev/null and b/egs/librispeech/WSASR/figures/ins.png differ
diff --git a/egs/librispeech/WSASR/figures/otc_emission.drawio.png b/egs/librispeech/WSASR/figures/otc_emission.drawio.png
new file mode 100644
index 000000000..6cea5531d
Binary files /dev/null and b/egs/librispeech/WSASR/figures/otc_emission.drawio.png differ
diff --git a/egs/librispeech/WSASR/figures/otc_g.png b/egs/librispeech/WSASR/figures/otc_g.png
new file mode 100644
index 000000000..ebad49180
Binary files /dev/null and b/egs/librispeech/WSASR/figures/otc_g.png differ
diff --git a/egs/librispeech/WSASR/figures/otc_training_graph.drawio.png b/egs/librispeech/WSASR/figures/otc_training_graph.drawio.png
new file mode 100644
index 000000000..8978158d8
Binary files /dev/null and b/egs/librispeech/WSASR/figures/otc_training_graph.drawio.png differ
diff --git a/egs/librispeech/WSASR/figures/sub.png b/egs/librispeech/WSASR/figures/sub.png
new file mode 100644
index 000000000..5674e9feb
Binary files /dev/null and b/egs/librispeech/WSASR/figures/sub.png differ
diff --git a/egs/librispeech/WSASR/local/compile_hlg.py b/egs/librispeech/WSASR/local/compile_hlg.py
new file mode 100755
index 000000000..63791f4cc
--- /dev/null
+++ b/egs/librispeech/WSASR/local/compile_hlg.py
@@ -0,0 +1,173 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input lang_dir and generates HLG from
+
+    - H, the ctc topology, built from tokens contained in lang_dir/lexicon.txt
+    - L, the lexicon, built from lang_dir/L_disambig.pt
+
+        Caution: We use a lexicon that contains disambiguation symbols
+
+    - G, the LM, built from data/lm/G_n_gram.fst.txt
+
+The generated HLG is saved in $lang_dir/HLG.pt
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lm",
+        type=str,
+        default="G_3_gram",
+        help="""Stem name for LM used in HLG compiling.
+        """,
+    )
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        help="""LM directory.
+        """,
+    )
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def compile_HLG(lm_dir: str, lang_dir: str, lm: str = "G_3_gram") -> k2.Fsa:
+    """
+    Args:
+      lang_dir:
+        The language directory, e.g., data/lang_phone or data/lang_bpe_5000.
+      lm:
+        The language stem base name.
+
+    Return:
+      An FSA representing HLG.
+    """
+    lexicon = Lexicon(lang_dir)
+    max_token_id = max(lexicon.tokens)
+    logging.info(f"Building ctc_topo. max_token_id: {max_token_id}")
+    H = k2.ctc_topo(max_token_id)
+    L = k2.Fsa.from_dict(torch.load(f"{lang_dir}/L_disambig.pt"))
+
+    if Path(f"{lm_dir}/{lm}.pt").is_file():
+        logging.info(f"Loading pre-compiled {lm}")
+        d = torch.load(f"{lm_dir}/{lm}.pt")
+        G = k2.Fsa.from_dict(d)
+    else:
+        logging.info(f"Loading {lm}.fst.txt")
+        with open(f"{lm_dir}/{lm}.fst.txt") as f:
+            G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+            torch.save(G.as_dict(), f"{lm_dir}/{lm}.pt")
+
+    first_token_disambig_id = lexicon.token_table["#0"]
+    first_word_disambig_id = lexicon.word_table["#0"]
+
+    L = k2.arc_sort(L)
+    G = k2.arc_sort(G)
+
+    logging.info("Intersecting L and G")
+    LG = k2.compose(L, G)
+    logging.info(f"LG shape: {LG.shape}")
+
+    logging.info("Connecting LG")
+    LG = k2.connect(LG)
+    logging.info(f"LG shape after k2.connect: {LG.shape}")
+
+    logging.info(type(LG.aux_labels))
+    logging.info("Determinizing LG")
+
+    LG = k2.determinize(LG)
+    logging.info(type(LG.aux_labels))
+
+    logging.info("Connecting LG after k2.determinize")
+    LG = k2.connect(LG)
+
+    logging.info("Removing disambiguation symbols on LG")
+
+    LG.labels[LG.labels >= first_token_disambig_id] = 0
+    # See https://github.com/k2-fsa/k2/issues/874
+    # for why we need to set LG.properties to None
+    LG.__dict__["_properties"] = None
+
+    assert isinstance(LG.aux_labels, k2.RaggedTensor)
+    LG.aux_labels.values[LG.aux_labels.values >= first_word_disambig_id] = 0
+
+    LG = k2.remove_epsilon(LG)
+    logging.info(f"LG shape after k2.remove_epsilon: {LG.shape}")
+
+    LG = k2.connect(LG)
+    LG.aux_labels = LG.aux_labels.remove_values_eq(0)
+
+    logging.info("Arc sorting LG")
+    LG = k2.arc_sort(LG)
+
+    logging.info("Composing H and LG")
+    # CAUTION: The name of the inner_labels is fixed
+    # to `tokens`. If you want to change it, please
+    # also change other places in icefall that are using
+    # it.
+    HLG = k2.compose(H, LG, inner_labels="tokens")
+
+    logging.info("Connecting LG")
+    HLG = k2.connect(HLG)
+
+    logging.info("Arc sorting LG")
+    HLG = k2.arc_sort(HLG)
+    logging.info(f"HLG.shape: {HLG.shape}")
+
+    return HLG
+
+
+def main():
+    args = get_args()
+    lm_dir = Path(args.lm_dir)
+    lang_dir = Path(args.lang_dir)
+
+    if (lang_dir / "HLG.pt").is_file():
+        logging.info(f"{lang_dir}/HLG.pt already exists - skipping")
+        return
+
+    logging.info(f"Processing {lang_dir}")
+
+    HLG = compile_HLG(lm_dir, lang_dir, args.lm)
+    logging.info(f"Saving HLG.pt to {lang_dir}")
+    torch.save(HLG.as_dict(), f"{lang_dir}/HLG.pt")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/WSASR/local/compute_fbank_librispeech.py b/egs/librispeech/WSASR/local/compute_fbank_librispeech.py
new file mode 100755
index 000000000..a387d54c9
--- /dev/null
+++ b/egs/librispeech/WSASR/local/compute_fbank_librispeech.py
@@ -0,0 +1,162 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the LibriSpeech dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import sentencepiece as spm
+import torch
+from filter_cuts import filter_cuts
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to the bpe.model. If not None, we will remove short and
+        long utterances before extracting features""",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=True,
+        help="""Perturb speed with factor 0.9 and 1.1 on train subset.""",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_librispeech(
+    bpe_model: Optional[str] = None,
+    dataset: Optional[str] = None,
+    perturb_speed: Optional[bool] = True,
+):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    if bpe_model:
+        logging.info(f"Loading {bpe_model}")
+        sp = spm.SentencePieceProcessor()
+        sp.load(bpe_model)
+
+    if dataset is None:
+        dataset_parts = (
+            "dev-clean",
+            "dev-other",
+            "test-clean",
+            "test-other",
+            "train-clean-100",
+        )
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    prefix = "librispeech"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+            if (output_dir / cuts_filename).is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+
+            if "train" in partition:
+                if bpe_model:
+                    cut_set = filter_cuts(cut_set, sp)
+                if perturb_speed:
+                    logging.info(f"Doing speed perturb")
+                    cut_set = (
+                        cut_set
+                        + cut_set.perturb_speed(0.9)
+                        + cut_set.perturb_speed(1.1)
+                    )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    compute_fbank_librispeech(
+        bpe_model=args.bpe_model,
+        dataset=args.dataset,
+        perturb_speed=args.perturb_speed,
+    )
diff --git a/egs/librispeech/WSASR/local/compute_ssl_librispeech.py b/egs/librispeech/WSASR/local/compute_ssl_librispeech.py
new file mode 100755
index 000000000..f405c468c
--- /dev/null
+++ b/egs/librispeech/WSASR/local/compute_ssl_librispeech.py
@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#              2023  Johns Hopkins University (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the LibriSpeech dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import S3PRLSSL, CutSet, NumpyFilesWriter, S3PRLSSLConfig
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_ssl_librispeech():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/ssl")
+    num_jobs = 1
+
+    dataset_parts = (
+        "dev-clean",
+        "dev-other",
+        "test-clean",
+        "test-other",
+        "train-clean-100",
+    )
+    prefix = "librispeech"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = S3PRLSSL(S3PRLSSLConfig(ssl_model="wav2vec2", device="cuda"))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+            if (output_dir / cuts_filename).is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                storage_type=NumpyFilesWriter,
+            )
+            cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_ssl_librispeech()
diff --git a/egs/librispeech/WSASR/local/filter_cuts.py b/egs/librispeech/WSASR/local/filter_cuts.py
new file mode 100644
index 000000000..fbcc9e24a
--- /dev/null
+++ b/egs/librispeech/WSASR/local/filter_cuts.py
@@ -0,0 +1,160 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script removes short and long utterances from a cutset.
+
+Caution:
+  You may need to tune the thresholds for your own dataset.
+
+Usage example:
+
+  python3 ./local/filter_cuts.py \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --in-cuts data/fbank/librispeech_cuts_test-clean.jsonl.gz \
+    --out-cuts data/fbank-filtered/librispeech_cuts_test-clean.jsonl.gz
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.cut import Cut
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=Path,
+        help="Path to the bpe.model",
+    )
+
+    parser.add_argument(
+        "--in-cuts",
+        type=Path,
+        help="Path to the input cutset",
+    )
+
+    parser.add_argument(
+        "--out-cuts",
+        type=Path,
+        help="Path to the output cutset",
+    )
+
+    return parser.parse_args()
+
+
+def filter_cuts(cut_set: CutSet, sp: spm.SentencePieceProcessor):
+    total = 0  # number of total utterances before removal
+    removed = 0  # number of removed utterances
+
+    def remove_short_and_long_utterances(c: Cut):
+        """Return False to exclude the input cut"""
+        nonlocal removed, total
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ./display_manifest_statistics.py
+        #
+        # You should use ./display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        total += 1
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            removed += 1
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./pruned_transducer_stateless2/conformer.py, the
+        # conv module uses the following expression
+        # for subsampling
+        if c.num_frames is None:
+            num_frames = c.duration * 100  # approximate
+        else:
+            num_frames = c.num_frames
+
+        T = ((num_frames - 1) // 2 - 1) // 2
+        # Note: for ./lstm_transducer_stateless/lstm.py, the formula is
+        #  T = ((num_frames - 3) // 2 - 1) // 2
+
+        # Note: for ./pruned_transducer_stateless7/zipformer.py, the formula is
+        # T = ((num_frames - 7) // 2 + 1) // 2
+
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            removed += 1
+            return False
+
+        return True
+
+    # We use to_eager() here so that we can print out the value of total
+    # and removed below.
+    ans = cut_set.filter(remove_short_and_long_utterances).to_eager()
+    ratio = removed / total * 100
+    logging.info(
+        f"Removed {removed} cuts from {total} cuts. {ratio:.3f}% data is removed."
+    )
+    return ans
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    if args.out_cuts.is_file():
+        logging.info(f"{args.out_cuts} already exists - skipping")
+        return
+
+    assert args.in_cuts.is_file(), f"{args.in_cuts} does not exist"
+    assert args.bpe_model.is_file(), f"{args.bpe_model} does not exist"
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(args.bpe_model))
+
+    cut_set = load_manifest_lazy(args.in_cuts)
+    assert isinstance(cut_set, CutSet)
+
+    cut_set = filter_cuts(cut_set, sp)
+    logging.info(f"Saving to {args.out_cuts}")
+    args.out_cuts.parent.mkdir(parents=True, exist_ok=True)
+    cut_set.to_file(args.out_cuts)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/WSASR/local/get_words_from_lexicon.py b/egs/librispeech/WSASR/local/get_words_from_lexicon.py
new file mode 100755
index 000000000..0cc740b36
--- /dev/null
+++ b/egs/librispeech/WSASR/local/get_words_from_lexicon.py
@@ -0,0 +1,48 @@
+#!/usr/bin/env python3
+
+import argparse
+from pathlib import Path
+
+from icefall.lexicon import read_lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain a file lexicon.txt.
+        Generated files by this script are saved into this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--otc-token",
+        type=str,
+        help="OTC token to be added to words.txt",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    otc_token = args.otc_token
+
+    lexicon = read_lexicon(lang_dir / "lexicon.txt")
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    words = ["<eps>"] + sorted_ans + [otc_token] + ["#0", "<s>", "</s>"]
+
+    words_file = lang_dir / "words.txt"
+    with open(words_file, "w") as wf:
+        for i, word in enumerate(words):
+            wf.write(f"{word} {i}\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/local/make_error_cutset.py b/egs/librispeech/WSASR/local/make_error_cutset.py
new file mode 100755
index 000000000..8463a380e
--- /dev/null
+++ b/egs/librispeech/WSASR/local/make_error_cutset.py
@@ -0,0 +1,175 @@
+#!/usr/bin/env python3
+
+# Copyright 2023 Johns Hopkins University (author: Dongji Gao)
+
+import argparse
+import random
+from pathlib import Path
+from typing import List
+
+from lhotse import CutSet, load_manifest
+from lhotse.cut.base import Cut
+
+from icefall.utils import str2bool
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    parser.add_argument(
+        "--input-cutset",
+        type=str,
+        help="Supervision manifest that contains verbatim transcript",
+    )
+
+    parser.add_argument(
+        "--words-file",
+        type=str,
+        help="words.txt file",
+    )
+
+    parser.add_argument(
+        "--otc-token",
+        type=str,
+        help="OTC token in words.txt",
+    )
+
+    parser.add_argument(
+        "--sub-error-rate",
+        type=float,
+        default=0.0,
+        help="Substitution error rate",
+    )
+
+    parser.add_argument(
+        "--ins-error-rate",
+        type=float,
+        default=0.0,
+        help="Insertion error rate",
+    )
+
+    parser.add_argument(
+        "--del-error-rate",
+        type=float,
+        default=0.0,
+        help="Deletion error rate",
+    )
+
+    parser.add_argument(
+        "--output-cutset",
+        type=str,
+        default="",
+        help="Supervision manifest that contains modified non-verbatim transcript",
+    )
+
+    parser.add_argument("--verbose", type=str2bool, help="show details of errors")
+    return parser.parse_args()
+
+
+def check_args(args):
+    total_error_rate = args.sub_error_rate + args.ins_error_rate + args.del_error_rate
+    assert args.sub_error_rate >= 0 and args.sub_error_rate <= 1.0
+    assert args.ins_error_rate >= 0 and args.sub_error_rate <= 1.0
+    assert args.del_error_rate >= 0 and args.sub_error_rate <= 1.0
+    assert total_error_rate <= 1.0
+
+
+def get_word_list(token_path: str) -> List:
+    word_list = []
+    with open(Path(token_path), "r") as tp:
+        for line in tp.readlines():
+            token = line.split()[0]
+            assert token not in word_list
+            word_list.append(token)
+    return word_list
+
+
+def modify_cut_text(
+    cut: Cut,
+    words_list: List,
+    non_words: List,
+    sub_ratio: float = 0.0,
+    ins_ratio: float = 0.0,
+    del_ratio: float = 0.0,
+):
+    text = cut.supervisions[0].text
+    text_list = text.split()
+
+    # We save the modified information of the original verbatim text for debugging
+    marked_verbatim_text_list = []
+    modified_text_list = []
+
+    del_index_set = set()
+    sub_index_set = set()
+    ins_index_set = set()
+
+    # We follow the order: deletion -> substitution -> insertion
+    for token in text_list:
+        marked_token = token
+        modified_token = token
+
+        prob = random.random()
+
+        if prob <= del_ratio:
+            marked_token = f"-{token}-"
+            modified_token = ""
+        elif prob <= del_ratio + sub_ratio + ins_ratio:
+            if prob <= del_ratio + sub_ratio:
+                marked_token = f"[{token}]"
+            else:
+                marked_verbatim_text_list.append(marked_token)
+                modified_text_list.append(modified_token)
+                marked_token = "[]"
+
+            # get new_token
+            while (
+                modified_token == token
+                or modified_token in non_words
+                or modified_token.startswith("#")
+            ):
+                modified_token = random.choice(words_list)
+
+        marked_verbatim_text_list.append(marked_token)
+        modified_text_list.append(modified_token)
+
+    marked_text = " ".join(marked_verbatim_text_list)
+    modified_text = " ".join(modified_text_list)
+
+    if not hasattr(cut.supervisions[0], "verbatim_text"):
+        cut.supervisions[0].verbatim_text = marked_text
+    cut.supervisions[0].text = modified_text
+
+    return cut
+
+
+def main():
+    args = get_args()
+    check_args(args)
+
+    otc_token = args.otc_token
+    non_words = set(("sil", "<UNK>", "<eps>"))
+    non_words.add(otc_token)
+
+    words_list = get_word_list(args.words_file)
+    cutset = load_manifest(Path(args.input_cutset))
+
+    cuts = []
+
+    for cut in cutset:
+        modified_cut = modify_cut_text(
+            cut=cut,
+            words_list=words_list,
+            non_words=non_words,
+            sub_ratio=args.sub_error_rate,
+            ins_ratio=args.ins_error_rate,
+            del_ratio=args.del_error_rate,
+        )
+        cuts.append(modified_cut)
+
+    output_cutset = CutSet.from_cuts(cuts)
+    output_cutset.to_file(args.output_cutset)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/local/prepare_lang.py b/egs/librispeech/WSASR/local/prepare_lang.py
new file mode 100755
index 000000000..d913756a1
--- /dev/null
+++ b/egs/librispeech/WSASR/local/prepare_lang.py
@@ -0,0 +1,413 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import argparse
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+from icefall.utils import str2bool
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain a file lexicon.txt.
+        Generated files by this script are saved into this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--debug",
+        type=str2bool,
+        default=False,
+        help="""True for debugging, which will generate
+        a visualization of the lexicon FST.
+
+        Caution: If your lexicon contains hundreds of thousands
+        of lines, please set it to False!
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    sil_token = token2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    lexicon_filename = lang_dir / "lexicon.txt"
+    sil_token = "SIL"
+    sil_prob = 0.5
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+    words = get_words(lexicon)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(lang_dir / "tokens.txt", token2id)
+    write_mapping(lang_dir / "words.txt", word2id)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+    if args.debug:
+        labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+        L.labels_sym = labels_sym
+        L.aux_labels_sym = aux_labels_sym
+        L.draw(f"{lang_dir / 'L.svg'}", title="L.pt")
+
+        L_disambig.labels_sym = labels_sym
+        L_disambig.aux_labels_sym = aux_labels_sym
+        L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/local/prepare_otc_lang_bpe.py b/egs/librispeech/WSASR/local/prepare_otc_lang_bpe.py
new file mode 100755
index 000000000..415bdff6f
--- /dev/null
+++ b/egs/librispeech/WSASR/local/prepare_otc_lang_bpe.py
@@ -0,0 +1,295 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#              2023  Johns Hopkins University (author: Dongji Gao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/bpe.model,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+from icefall.utils import str2bool
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def generate_otc_lexicon(
+    model_file: str,
+    words: List[str],
+    oov: str,
+    otc_token: str,
+) -> Tuple[Lexicon, Dict[str, int]]:
+    """Generate a lexicon from a BPE model.
+
+    Args:
+      model_file:
+        Path to a sentencepiece model.
+      words:
+        A list of strings representing words.
+      oov:
+        The out of vocabulary word in lexicon.
+      otc_token:
+        The OTC token in lexicon.
+    Returns:
+      Return a tuple with two elements:
+        - A dict whose keys are words and values are the corresponding
+          word pieces.
+        - A dict representing the token symbol, mapping from tokens to IDs.
+    """
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(model_file))
+
+    # Convert word to word piece IDs instead of word piece strings
+    # to avoid OOV tokens.
+    words_pieces_ids: List[List[int]] = sp.encode(words, out_type=int)
+
+    # Now convert word piece IDs back to word piece strings.
+    words_pieces: List[List[str]] = [sp.id_to_piece(ids) for ids in words_pieces_ids]
+
+    lexicon = []
+    for word, pieces in zip(words, words_pieces):
+        lexicon.append((word, pieces))
+
+    lexicon.append((oov, ["▁", sp.id_to_piece(sp.unk_id())]))
+    token2id: Dict[str, int] = {sp.id_to_piece(i): i for i in range(sp.vocab_size())}
+
+    # Add OTC token to the last.
+    lexicon.append((otc_token, [f"▁{otc_token}"]))
+    otc_token_index = len(token2id)
+    token2id[f"▁{otc_token}"] = otc_token_index
+
+    return lexicon, token2id
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain the bpe.model and words.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--oov",
+        type=str,
+        default="<UNK>",
+        help="The out of vocabulary word in lexicon.",
+    )
+
+    parser.add_argument(
+        "--otc-token",
+        type=str,
+        default="<star>",
+        help="The OTC token in lexicon.",
+    )
+
+    parser.add_argument(
+        "--debug",
+        type=str2bool,
+        default=False,
+        help="""True for debugging, which will generate
+        a visualization of the lexicon FST.
+
+        Caution: If your lexicon contains hundreds of thousands
+        of lines, please set it to False!
+
+        See "test/test_bpe_lexicon.py" for usage.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    model_file = lang_dir / "bpe.model"
+    otc_token = args.otc_token
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = [
+        "<eps>",
+        "!SIL",
+        "<SPOKEN_NOISE>",
+        args.oov,
+        otc_token,
+        "#0",
+        "<s>",
+        "</s>",
+    ]
+
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    lexicon, token_sym_table = generate_otc_lexicon(
+        model_file, words, args.oov, otc_token
+    )
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+    if args.debug:
+        labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+        L.labels_sym = labels_sym
+        L.aux_labels_sym = aux_labels_sym
+        L.draw(f"{lang_dir / 'L.svg'}", title="L.pt")
+
+        L_disambig.labels_sym = labels_sym
+        L_disambig.aux_labels_sym = aux_labels_sym
+        L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/local/train_bpe_model.py b/egs/librispeech/WSASR/local/train_bpe_model.py
new file mode 100755
index 000000000..43142aee4
--- /dev/null
+++ b/egs/librispeech/WSASR/local/train_bpe_model.py
@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# You can install sentencepiece via:
+#
+#  pip install sentencepiece
+#
+# Due to an issue reported in
+# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030
+#
+# Please install a version >=0.1.96
+
+import argparse
+import shutil
+from pathlib import Path
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        The generated bpe.model is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="Training transcript.",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        help="Vocabulary size for BPE training",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    vocab_size = args.vocab_size
+    lang_dir = Path(args.lang_dir)
+
+    model_type = "unigram"
+
+    model_prefix = f"{lang_dir}/{model_type}_{vocab_size}"
+    train_text = args.transcript
+    character_coverage = 1.0
+    input_sentence_size = 100000000
+
+    user_defined_symbols = ["<blk>", "<sos/eos>"]
+    unk_id = len(user_defined_symbols)
+    # Note: unk_id is fixed to 2.
+    # If you change it, you should also change other
+    # places that are using it.
+
+    model_file = Path(model_prefix + ".model")
+    if not model_file.is_file():
+        spm.SentencePieceTrainer.train(
+            input=train_text,
+            vocab_size=vocab_size,
+            model_type=model_type,
+            model_prefix=model_prefix,
+            input_sentence_size=input_sentence_size,
+            character_coverage=character_coverage,
+            user_defined_symbols=user_defined_symbols,
+            unk_id=unk_id,
+            bos_id=-1,
+            eos_id=-1,
+        )
+    else:
+        print(f"{model_file} exists - skipping")
+        return
+
+    shutil.copyfile(model_file, f"{lang_dir}/bpe.model")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/local/validate_bpe_lexicon.py b/egs/librispeech/WSASR/local/validate_bpe_lexicon.py
new file mode 100755
index 000000000..16a489c11
--- /dev/null
+++ b/egs/librispeech/WSASR/local/validate_bpe_lexicon.py
@@ -0,0 +1,85 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script checks that there are no OOV tokens in the BPE-based lexicon.
+
+Usage example:
+
+    python3 ./local/validate_bpe_lexicon.py \
+            --lexicon /path/to/lexicon.txt \
+            --bpe-model /path/to/bpe.model
+"""
+
+import argparse
+from pathlib import Path
+from typing import List, Tuple
+
+import sentencepiece as spm
+
+from icefall.lexicon import read_lexicon
+
+# Map word to word pieces
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--lexicon",
+        required=True,
+        type=Path,
+        help="Path to lexicon.txt",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        required=True,
+        type=Path,
+        help="Path to bpe.model",
+    )
+
+    parser.add_argument(
+        "--otc-token",
+        required=True,
+        type=str,
+        help="OTC token",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    assert args.lexicon.is_file(), args.lexicon
+    assert args.bpe_model.is_file(), args.bpe_model
+
+    lexicon = read_lexicon(args.lexicon)
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(args.bpe_model))
+
+    word_pieces = set(sp.id_to_piece(list(range(sp.vocab_size()))))
+    word_pieces.add(f"▁{args.otc_token}")
+    for word, pieces in lexicon:
+        for p in pieces:
+            if p not in word_pieces:
+                raise ValueError(f"The word {word} contains an OOV token {p}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/librispeech/WSASR/local/validate_manifest.py b/egs/librispeech/WSASR/local/validate_manifest.py
new file mode 100755
index 000000000..f620b91ea
--- /dev/null
+++ b/egs/librispeech/WSASR/local/validate_manifest.py
@@ -0,0 +1,92 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script checks the following assumptions of the generated manifest:
+
+- Single supervision per cut
+- Supervision time bounds are within cut time bounds
+
+We will add more checks later if needed.
+
+Usage example:
+
+    python3 ./local/validate_manifest.py \
+            ./data/fbank/librispeech_cuts_train-clean-100.jsonl.gz
+
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.cut import Cut
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "manifest",
+        type=Path,
+        help="Path to the manifest file",
+    )
+
+    return parser.parse_args()
+
+
+def validate_one_supervision_per_cut(c: Cut):
+    if len(c.supervisions) != 1:
+        raise ValueError(f"{c.id} has {len(c.supervisions)} supervisions")
+
+
+def validate_supervision_and_cut_time_bounds(c: Cut):
+    s = c.supervisions[0]
+    if s.start < c.start:
+        raise ValueError(
+            f"{c.id}: Supervision start time {s.start} is less "
+            f"than cut start time {c.start}"
+        )
+
+    if s.end > c.end:
+        raise ValueError(
+            f"{c.id}: Supervision end time {s.end} is larger "
+            f"than cut end time {c.end}"
+        )
+
+
+def main():
+    args = get_args()
+
+    manifest = args.manifest
+    logging.info(f"Validating {manifest}")
+
+    assert manifest.is_file(), f"{manifest} does not exist"
+    cut_set = load_manifest_lazy(manifest)
+    assert isinstance(cut_set, CutSet)
+
+    for c in cut_set:
+        validate_one_supervision_per_cut(c)
+        validate_supervision_and_cut_time_bounds(c)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/WSASR/prepare.sh b/egs/librispeech/WSASR/prepare.sh
new file mode 100755
index 000000000..f6a922fde
--- /dev/null
+++ b/egs/librispeech/WSASR/prepare.sh
@@ -0,0 +1,233 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/LibriSpeech
+#      You can find BOOKS.TXT, test-clean, train-clean-360, etc, inside it.
+#      You can download them from https://www.openslr.org/12
+#
+#  - $dl_dir/lm
+#      This directory contains the following files downloaded from
+#       http://www.openslr.org/resources/11
+#
+#        - 3-gram.pruned.1e-7.arpa.gz
+#        - 3-gram.pruned.1e-7.arpa
+#        - 4-gram.arpa.gz
+#        - 4-gram.arpa
+#        - librispeech-vocab.txt
+#        - librispeech-lexicon.txt
+#        - librispeech-lm-norm.txt.gz
+#
+otc_token="<star>"
+feature_type="ssl"
+
+dl_dir=$PWD/download
+manifests_dir="data/manifests"
+feature_dir="data/${feature_type}"
+lang_dir="data/lang"
+lm_dir="data/lm"
+
+perturb_speed=false
+
+# ssl or fbank
+
+. ./cmd.sh
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  200
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: ${dl_dir}"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: Download LM"
+  mkdir -p ${dl_dir}/lm
+  if [ ! -e ${dl_dir}/lm/.done ]; then
+    ./local/download_lm.py --out-dir=${dl_dir}/lm
+    touch ${dl_dir}/lm/.done
+  fi
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/LibriSpeech,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/LibriSpeech $dl_dir/LibriSpeech
+  #
+  if [ ! -d $dl_dir/LibriSpeech/train-clean-100 ]; then
+    lhotse download librispeech --full ${dl_dir}
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare LibriSpeech manifest"
+  # We assume that you have downloaded the LibriSpeech corpus
+  # to $dl_dir/LibriSpeech
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.librispeech.done ]; then
+    lhotse prepare librispeech -j ${nj} \
+      -p dev-clean \
+      -p dev-other \
+      -p test-clean \
+      -p test-other \
+      -p train-clean-100 "${dl_dir}/LibriSpeech" "${manifests_dir}"
+    touch data/manifests/.librispeech.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Compute ${feature_type} feature for librispeech (train-clean-100)"
+  mkdir -p "${feature_dir}"  
+  if [ ! -e "${feature_dir}/.librispeech.done" ]; then
+    if [ "${feature_type}" = ssl ]; then
+      ./local/compute_ssl_librispeech.py
+    elif [ "${feature_type}" = fbank ]; then
+      ./local/compute_fbank_librispeech.py --perturb-speed ${perturb_speed}
+    else
+      log "Error: not supported --feature-type '${feature_type}'" 
+      exit 2
+    fi
+
+    touch "${feature_dir}.librispeech.done"
+  fi
+
+  if [ ! -e "${feature_dir}/.librispeech-validated.done" ]; then
+    log "Validating data/ssl for LibriSpeech"
+    parts=(
+      train-clean-100
+      test-clean
+      test-other
+      dev-clean
+      dev-other
+    )
+    for part in ${parts[@]}; do
+      python3 ./local/validate_manifest.py \
+        "${feature_dir}/librispeech_cuts_${part}.jsonl.gz"
+    done
+    touch "${feature_dir}/.librispeech-validated.done"
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare words.txt"
+  mkdir -p ${lang_dir}
+
+  (echo '!SIL SIL'; echo '<SPOKEN_NOISE> SPN'; echo '<UNK> SPN'; ) |
+    cat - $dl_dir/lm/librispeech-lexicon.txt |
+    sort | uniq > ${lang_dir}/lexicon.txt
+
+  local/get_words_from_lexicon.py \
+    --lang-dir ${lang_dir} \
+    --otc-token ${otc_token}
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    bpe_lang_dir="data/lang_bpe_${vocab_size}"
+    mkdir -p "${bpe_lang_dir}"
+    # We reuse words.txt from phone based lexicon
+    # so that the two can share G.pt later.
+    cp "${lang_dir}/words.txt" "${bpe_lang_dir}"
+
+    if [ ! -f "${bpe_lang_dir}/transcript_words.txt" ]; then
+      log "Generate data for BPE training"
+      files=$(
+        find "$dl_dir/LibriSpeech/train-clean-100" -name "*.trans.txt"
+        find "$dl_dir/LibriSpeech/train-clean-360" -name "*.trans.txt"
+        find "$dl_dir/LibriSpeech/train-other-500" -name "*.trans.txt"
+      )
+      for f in ${files[@]}; do
+        cat $f | cut -d " " -f 2-
+      done > "${bpe_lang_dir}/transcript_words.txt"
+    fi
+
+    if [ ! -f ${bpe_lang_dir}/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir ${bpe_lang_dir} \
+        --vocab-size ${vocab_size} \
+        --transcript ${bpe_lang_dir}/transcript_words.txt 
+    fi
+
+    if [ ! -f ${bpe_lang_dir}/L_disambig.pt ]; then
+      ./local/prepare_otc_lang_bpe.py \
+        --lang-dir "${bpe_lang_dir}" \
+        --otc-token "${otc_token}"
+
+      log "Validating ${bpe_lang_dir}/lexicon.txt"
+      ./local/validate_bpe_lexicon.py \
+        --lexicon ${bpe_lang_dir}/lexicon.txt \
+        --bpe-model ${bpe_lang_dir}/bpe.model \
+        --otc-token "${otc_token}"
+    fi
+  done
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare G"
+  # We assume you have install kaldilm, if not, please install
+  # it using: pip install kaldilm
+
+  mkdir -p "${lm_dir}"
+  if [ ! -f ${lm_dir}/G_3_gram.fst.txt ]; then
+    # It is used in building HLG
+    python3 -m kaldilm \
+      --read-symbol-table="${lang_dir}/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=3 \
+      ${dl_dir}/lm/3-gram.pruned.1e-7.arpa > ${lm_dir}/G_3_gram.fst.txt
+  fi
+
+  if [ ! -f ${lm_dir}/G_4_gram.fst.txt ]; then
+    # It is used for LM rescoring
+    python3 -m kaldilm \
+      --read-symbol-table="${lang_dir}/words.txt" \
+      --disambig-symbol='#0' \
+      --max-order=4 \
+      ${dl_dir}/lm/4-gram.arpa > ${lm_dir}/G_4_gram.fst.txt
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Compile HLG"
+  # Note If ./local/compile_hlg.py throws OOM,
+  # please switch to the following command
+  #
+  # ./local/compile_hlg_using_openfst.py --lang-dir data/lang_phone
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    bpe_lang_dir="data/lang_bpe_${vocab_size}"
+    echo "LM DIR: ${lm_dir}"
+    ./local/compile_hlg.py \
+        --lm-dir "${lm_dir}" \
+        --lang-dir "${bpe_lang_dir}"
+  done
+fi
diff --git a/egs/mgb2/ASR/conformer_ctc/asr_datamodule.py b/egs/mgb2/ASR/conformer_ctc/asr_datamodule.py
index 442ff85c2..7753d1674 100644
--- a/egs/mgb2/ASR/conformer_ctc/asr_datamodule.py
+++ b/egs/mgb2/ASR/conformer_ctc/asr_datamodule.py
@@ -182,7 +182,6 @@ class MGB2AsrDataModule:
         cuts_train: CutSet,
         sampler_state_dict: Optional[Dict[str, Any]] = None,
     ) -> DataLoader:
-
         transforms = []
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
@@ -190,7 +189,7 @@ class MGB2AsrDataModule:
             cuts_musan = load_manifest(self.args.manifest_dir / "cuts_musan.jsonl.gz")
 
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/mgb2/ASR/pruned_transducer_stateless5/train.py b/egs/mgb2/ASR/pruned_transducer_stateless5/train.py
index a68702776..48468cfbd 100755
--- a/egs/mgb2/ASR/pruned_transducer_stateless5/train.py
+++ b/egs/mgb2/ASR/pruned_transducer_stateless5/train.py
@@ -746,7 +746,6 @@ def train_one_epoch(
     tot_loss = MetricsTracker()
 
     for batch_idx, batch in enumerate(train_dl):
-
         if batch["inputs"].shape[0] == len(batch["supervisions"]["text"]):
             params.batch_idx_train += 1
             batch_size = len(batch["supervisions"]["text"])
@@ -966,7 +965,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
@@ -1019,7 +1018,6 @@ def run(rank, world_size, args):
         scaler.load_state_dict(checkpoints["grad_scaler"])
 
     for epoch in range(params.start_epoch, params.num_epochs + 1):
-
         scheduler.step_epoch(epoch - 1)
         fix_random_seed(params.seed + epoch - 1)
         train_dl.sampler.set_epoch(epoch - 1)
@@ -1118,7 +1116,6 @@ def scan_pessimistic_batches_for_oom(
             # (i.e. are not remembered by the decaying-average in adam), because
             # we want to avoid these params being subject to shrinkage in adam.
             with torch.cuda.amp.autocast(enabled=params.use_fp16):
-
                 loss, _, _ = compute_loss(
                     params=params,
                     model=model,
diff --git a/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py b/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py
index 3d58ebf3a..02cfa1346 100644
--- a/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py
+++ b/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py
@@ -219,7 +219,7 @@ class AsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/multi_zh-hans/ASR/zipformer/train.py b/egs/multi_zh-hans/ASR/zipformer/train.py
index 4f2d728be..c1bbd2ee8 100755
--- a/egs/multi_zh-hans/ASR/zipformer/train.py
+++ b/egs/multi_zh-hans/ASR/zipformer/train.py
@@ -1164,7 +1164,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
index d94a92503..cf70fc0f8 100644
--- a/egs/spgispeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
@@ -182,7 +182,7 @@ class SPGISpeechAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
@@ -261,7 +261,6 @@ class SPGISpeechAsrDataModule:
         return train_dl
 
     def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
-
         transforms = []
         if self.args.concatenate_cuts:
             transforms = [
diff --git a/egs/swbd/ASR/.gitignore b/egs/swbd/ASR/.gitignore
new file mode 100644
index 000000000..11d674922
--- /dev/null
+++ b/egs/swbd/ASR/.gitignore
@@ -0,0 +1,2 @@
+switchboard_word_alignments.tar.gz
+./swb_ms98_transcriptions/
diff --git a/egs/swbd/ASR/README.md b/egs/swbd/ASR/README.md
new file mode 100644
index 000000000..13b27815a
--- /dev/null
+++ b/egs/swbd/ASR/README.md
@@ -0,0 +1,25 @@
+# Switchboard
+
+The Switchboard-1 Telephone Speech Corpus (LDC97S62) consists of approximately 260 hours of speech and was originally collected by Texas Instruments in 1990-1, under DARPA sponsorship. The first release of the corpus was published by NIST and distributed by the LDC in 1992-3. Since that release, a number of corrections have been made to the data files as presented on the original CD-ROM set and all copies of the first pressing have been distributed.
+
+Switchboard is a collection of about 2,400 two-sided telephone conversations among 543 speakers (302 male, 241 female) from all areas of the United States. A computer-driven robot operator system handled the calls, giving the caller appropriate recorded prompts, selecting and dialing another person (the callee) to take part in a conversation, introducing a topic for discussion and recording the speech from the two subjects into separate channels until the conversation was finished. About 70 topics were provided, of which about 50 were used frequently. Selection of topics and callees was constrained so that: (1) no two speakers would converse together more than once and (2) no one spoke more than once on a given topic.
+
+(The above introduction is from the [LDC Switchboard-1 Release 2 webpage](https://catalog.ldc.upenn.edu/LDC97S62).)
+
+
+## Performance Record
+|                                |  eval2000  |  rt03  |
+|--------------------------------|------------|--------|
+|         `conformer_ctc`        |    33.37   |  35.06 |
+
+See [RESULTS](/egs/swbd/ASR/RESULTS.md) for details.
+
+## Credit
+
+The training script for `conformer_ctc` comes from the LibriSpeech `conformer_ctc` recipe in icefall.
+
+A lot of the scripts for data processing are from the first-gen Kaldi and the ESPNet project, tailored by myself to incorporate with Lhotse and Icefall.
+
+Some of the scripts for text normalization are from stale pull requests of [Piotr Żelasko](https://github.com/pzelasko) and [Nagendra Goel](https://github.com/ngoel17).
+
+The `sclite_scoring.py` is from the GigaSpeech recipe for post processing and glm-like scoring, which is definitely not an elegant stuff to do.
diff --git a/egs/swbd/ASR/RESULTS.md b/egs/swbd/ASR/RESULTS.md
new file mode 100644
index 000000000..f3a22c444
--- /dev/null
+++ b/egs/swbd/ASR/RESULTS.md
@@ -0,0 +1,113 @@
+## Results
+### Switchboard BPE training results (Conformer-CTC)
+
+#### 2023-09-04
+
+The best WER, as of 2023-09-04, for the Switchboard is below
+
+Results using attention decoder are given as:
+
+|                                |  eval2000-swbd  |  eval2000-callhome  | eval2000-avg |
+|--------------------------------|-----------------|---------------------|--------------|
+|         `conformer_ctc`        |      9.48       |         17.73       |    13.67     | 
+
+Decoding results and models can be found here:
+https://huggingface.co/zrjin/icefall-asr-swbd-conformer-ctc-2023-8-26
+#### 2023-06-27
+
+The best WER, as of 2023-06-27, for the Switchboard is below
+
+Results using HLG decoding + n-gram LM rescoring + attention decoder rescoring:
+
+|                                |  eval2000  |  rt03  |
+|--------------------------------|------------|--------|
+|         `conformer_ctc`        |    30.80   |  32.29 |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+
+##### eval2000
+
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+|      0.9       |       1.1       |
+
+##### rt03
+
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+|      0.9       |       1.9       |
+
+To reproduce the above result, use the following commands for training:
+
+```bash
+cd egs/swbd/ASR
+./prepare.sh
+export CUDA_VISIBLE_DEVICES="0,1"
+./conformer_ctc/train.py \
+  --max-duration 120 \
+  --num-workers 8 \
+  --enable-musan False \
+  --world-size 2 \
+  --num-epochs 100
+```
+
+and the following command for decoding:
+
+```bash
+./conformer_ctc/decode.py \
+  --epoch 99 \
+  --avg 10 \
+  --max-duration 50
+```
+
+#### 2023-06-26
+
+The best WER, as of 2023-06-26, for the Switchboard is below
+
+Results using HLG decoding + n-gram LM rescoring + attention decoder rescoring:
+
+|                                |  eval2000  |  rt03  |
+|--------------------------------|------------|--------|
+|         `conformer_ctc`        |    33.37   |  35.06 |
+
+Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
+
+##### eval2000
+
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+|      0.3       |       2.5       |
+
+##### rt03
+
+| ngram_lm_scale | attention_scale |
+|----------------|-----------------|
+|      0.7       |       1.3       |
+
+To reproduce the above result, use the following commands for training:
+
+```bash
+cd egs/swbd/ASR
+./prepare.sh
+export CUDA_VISIBLE_DEVICES="0,1"
+./conformer_ctc/train.py \
+  --max-duration 120 \
+  --num-workers 8 \
+  --enable-musan False \
+  --world-size 2 \
+```
+
+and the following command for decoding:
+
+```bash
+./conformer_ctc/decode.py \
+  --epoch 55 \
+  --avg 1 \
+  --max-duration 50
+```
+
+For your reference, the nbest oracle WERs are:
+
+|                                |  eval2000  |  rt03  |
+|--------------------------------|------------|--------|
+|         `conformer_ctc`        |    25.64   |  26.84 |
diff --git a/egs/swbd/ASR/conformer_ctc/__init__.py b/egs/swbd/ASR/conformer_ctc/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/egs/swbd/ASR/conformer_ctc/asr_datamodule.py b/egs/swbd/ASR/conformer_ctc/asr_datamodule.py
new file mode 100644
index 000000000..ce8634a1d
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/asr_datamodule.py
@@ -0,0 +1,416 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang Luo)
+# Modified by Zengrui Jin for the SwitchBoard corpus
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SimpleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class SwitchBoardAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train dataloader,
+    but there can be multiple test dataloaders (e.g. SwitchBoard rt03
+    and eval2000).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=50,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+                buffer_size=50000,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else eval(self.args.input_strategy)(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_all_cuts(self) -> CutSet:
+        logging.info("SwitchBoard: About to get train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "swbd_cuts_all.jsonl.gz"
+        ).subset(last=166844)
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("SwitchBoard: About to get dev cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "swbd_cuts_all.jsonl.gz"
+        ).subset(first=300)
+
+    @lru_cache()
+    def test_eval2000_cuts(self) -> CutSet:
+        logging.info("SwitchBoard: About to get eval2000 cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "eval2000" / "eval2000_cuts_all.jsonl.gz"
+        )
+
+    @lru_cache()
+    def test_rt03_cuts(self) -> CutSet:
+        logging.info("SwitchBoard: About to get rt03 cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "swbd_cuts_rt03.jsonl.gz")
diff --git a/egs/swbd/ASR/conformer_ctc/conformer.py b/egs/swbd/ASR/conformer_ctc/conformer.py
new file mode 120000
index 000000000..d1f4209d7
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/conformer.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/conformer_ctc/decode.py b/egs/swbd/ASR/conformer_ctc/decode.py
new file mode 100755
index 000000000..2bbade374
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/decode.py
@@ -0,0 +1,853 @@
+#!/usr/bin/env python3
+# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, Fangjun Kuang)
+# Modified by Zengrui Jin for the SwitchBoard corpus
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import SwitchBoardAsrDataModule
+from conformer import Conformer
+
+from sclite_scoring import asr_text_post_processing
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.decode import (
+    get_lattice,
+    nbest_decoding,
+    nbest_oracle,
+    one_best_decoding,
+    rescore_with_attention_decoder,
+    rescore_with_n_best_list,
+    rescore_with_rnn_lm,
+    rescore_with_whole_lattice,
+)
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.rnn_lm.model import RnnLmModel
+from icefall.utils import (
+    AttributeDict,
+    get_texts,
+    load_averaged_model,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=98,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=55,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="attention-decoder",
+        help="""Decoding method.
+        Supported values are:
+            - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece
+              model, i.e., lang_dir/bpe.model, to convert word pieces to words.
+              It needs neither a lexicon nor an n-gram LM.
+            - (1) 1best. Extract the best path from the decoding lattice as the
+              decoding result.
+            - (2) nbest. Extract n paths from the decoding lattice; the path
+              with the highest score is the decoding result.
+            - (3) nbest-rescoring. Extract n paths from the decoding lattice,
+              rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
+              the highest score is the decoding result.
+            - (4) whole-lattice-rescoring. Rescore the decoding lattice with an
+              n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
+              is the decoding result.
+            - (5) attention-decoder. Extract n paths from the LM rescored
+              lattice, the path with the highest score is the decoding result.
+            - (6) rnn-lm. Rescoring with attention-decoder and RNN LM. We assume
+              you have trained an RNN LM using ./rnn_lm/train.py
+            - (7) nbest-oracle. Its WER is the lower bound of any n-best
+              rescoring method can achieve. Useful for debugging n-best
+              rescoring method.
+        """,
+    )
+
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=100,
+        help="""Number of paths for n-best based decoding method.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""The scale to be applied to `lattice.scores`.
+        It's needed if you use any kinds of n-best based rescoring.
+        Used only when "method" is one of the following values:
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
+        A smaller value results in more unique paths.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--lm-dir",
+        type=str,
+        default="data/lm",
+        help="""The n-gram LM dir.
+        It should contain either G_4_gram.pt or G_4_gram.fst.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""Used only when --method is rnn-lm.
+        It specifies the path to RNN LM exp dir.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-epoch",
+        type=int,
+        default=7,
+        help="""Used only when --method is rnn-lm.
+        It specifies the checkpoint to use.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-avg",
+        type=int,
+        default=2,
+        help="""Used only when --method is rnn-lm.
+        It specifies the number of checkpoints to average.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-num-layers",
+        type=int,
+        default=4,
+        help="Number of RNN layers the model",
+    )
+    parser.add_argument(
+        "--rnn-lm-tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            # parameters for conformer
+            "subsampling_factor": 4,
+            "vgg_frontend": False,
+            "use_feat_batchnorm": True,
+            "feature_dim": 80,
+            "nhead": 8,
+            "attention_dim": 512,
+            "num_decoder_layers": 6,
+            # parameters for decoding
+            "search_beam": 20,
+            "output_beam": 8,
+            "min_active_states": 30,
+            "max_active_states": 10000,
+            "use_double_scores": True,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def post_processing(
+    results: List[Tuple[str, List[str], List[str]]],
+) -> List[Tuple[str, List[str], List[str]]]:
+    new_results = []
+    for key, ref, hyp in results:
+        new_ref = asr_text_post_processing(" ".join(ref)).split()
+        new_hyp = asr_text_post_processing(" ".join(hyp)).split()
+        new_results.append((key, new_ref, new_hyp))
+    return new_results
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    batch: dict,
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if no rescoring is used, the key is the string `no_rescore`.
+               If LM rescoring is used, the key is the string `lm_scale_xxx`,
+               where `xxx` is the value of `lm_scale`. An example key is
+               `lm_scale_0.7`
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+
+        - params.method is "1best", it uses 1best decoding without LM rescoring.
+        - params.method is "nbest", it uses nbest decoding without LM rescoring.
+        - params.method is "nbest-rescoring", it uses nbest LM rescoring.
+        - params.method is "whole-lattice-rescoring", it uses whole lattice LM
+          rescoring.
+
+      model:
+        The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      word_table:
+        The word symbol table.
+      sos_id:
+        The token ID of the SOS.
+      eos_id:
+        The token ID of the EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict. Note: If it decodes to nothing, then return None.
+    """
+    if HLG is not None:
+        device = HLG.device
+    else:
+        device = H.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+
+    nnet_output, memory, memory_key_padding_mask = model(feature, supervisions)
+    # nnet_output is (N, T, C)
+
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            supervisions["start_frame"] // params.subsampling_factor,
+            supervisions["num_frames"] // params.subsampling_factor,
+        ),
+        1,
+    ).to(torch.int32)
+
+    if H is None:
+        assert HLG is not None
+        decoding_graph = HLG
+    else:
+        assert HLG is None
+        assert bpe_model is not None
+        decoding_graph = H
+
+    lattice = get_lattice(
+        nnet_output=nnet_output,
+        decoding_graph=decoding_graph,
+        supervision_segments=supervision_segments,
+        search_beam=params.search_beam,
+        output_beam=params.output_beam,
+        min_active_states=params.min_active_states,
+        max_active_states=params.max_active_states,
+        subsampling_factor=params.subsampling_factor,
+    )
+
+    if params.method == "ctc-decoding":
+        best_path = one_best_decoding(
+            lattice=lattice, use_double_scores=params.use_double_scores
+        )
+        # Note: `best_path.aux_labels` contains token IDs, not word IDs
+        # since we are using H, not HLG here.
+        #
+        # token_ids is a lit-of-list of IDs
+        token_ids = get_texts(best_path)
+
+        # hyps is a list of str, e.g., ['xxx yyy zzz', ...]
+        hyps = bpe_model.decode(token_ids)
+
+        # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
+        hyps = [s.split() for s in hyps]
+        key = "ctc-decoding"
+        return {key: hyps}
+
+    if params.method == "nbest-oracle":
+        # Note: You can also pass rescored lattices to it.
+        # We choose the HLG decoded lattice for speed reasons
+        # as HLG decoding is faster and the oracle WER
+        # is only slightly worse than that of rescored lattices.
+        best_path = nbest_oracle(
+            lattice=lattice,
+            num_paths=params.num_paths,
+            ref_texts=supervisions["text"],
+            word_table=word_table,
+            nbest_scale=params.nbest_scale,
+            oov="<unk>",
+        )
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}"  # noqa
+        return {key: hyps}
+
+    if params.method in ["1best", "nbest"]:
+        if params.method == "1best":
+            best_path = one_best_decoding(
+                lattice=lattice, use_double_scores=params.use_double_scores
+            )
+            key = "no_rescore"
+        else:
+            best_path = nbest_decoding(
+                lattice=lattice,
+                num_paths=params.num_paths,
+                use_double_scores=params.use_double_scores,
+                nbest_scale=params.nbest_scale,
+            )
+            key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}"  # noqa
+
+        hyps = get_texts(best_path)
+        hyps = [[word_table[i] for i in ids] for ids in hyps]
+        return {key: hyps}
+
+    assert params.method in [
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ]
+
+    lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
+    lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
+    lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
+
+    if params.method == "nbest-rescoring":
+        best_path_dict = rescore_with_n_best_list(
+            lattice=lattice,
+            G=G,
+            num_paths=params.num_paths,
+            lm_scale_list=lm_scale_list,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "whole-lattice-rescoring":
+        best_path_dict = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=lm_scale_list,
+        )
+    elif params.method == "attention-decoder":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_attention_decoder(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            nbest_scale=params.nbest_scale,
+        )
+    elif params.method == "rnn-lm":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_rnn_lm(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            rnn_lm_model=rnn_lm_model,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            blank_id=0,
+            nbest_scale=params.nbest_scale,
+        )
+    else:
+        assert False, f"Unsupported decoding method: {params.method}"
+
+    ans = dict()
+    if best_path_dict is not None:
+        for lm_scale_str, best_path in best_path_dict.items():
+            hyps = get_texts(best_path)
+            hyps = [[word_table[i] for i in ids] for ids in hyps]
+            ans[lm_scale_str] = hyps
+    else:
+        ans = None
+    return ans
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
+    HLG: Optional[k2.Fsa],
+    H: Optional[k2.Fsa],
+    bpe_model: Optional[spm.SentencePieceProcessor],
+    word_table: k2.SymbolTable,
+    sos_id: int,
+    eos_id: int,
+    G: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
+      HLG:
+        The decoding graph. Used only when params.method is NOT ctc-decoding.
+      H:
+        The ctc topo. Used only when params.method is ctc-decoding.
+      bpe_model:
+        The BPE model. Used only when params.method is ctc-decoding.
+      word_table:
+        It is the word symbol table.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      G:
+        An LM. It is not None when params.method is "nbest-rescoring"
+        or "whole-lattice-rescoring". In general, the G in HLG
+        is a 3-gram LM, while this G is a 4-gram LM.
+    Returns:
+      Return a dict, whose key may be "no-rescore" if no LM rescoring
+      is used, or it may be "lm_scale_0.7" if LM rescoring is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            batch=batch,
+            word_table=word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        if hyps_dict is not None:
+            for lm_scale, hyps in hyps_dict.items():
+                this_batch = []
+                assert len(hyps) == len(texts)
+                for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                    ref_words = ref_text.split()
+                    this_batch.append((cut_id, ref_words, hyp_words))
+
+                results[lm_scale].extend(this_batch)
+        else:
+            assert len(results) > 0, "It should not decode to empty in the first batch!"
+            this_batch = []
+            hyp_words = []
+            for ref_text in texts:
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            for lm_scale in results.keys():
+                results[lm_scale].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 100 == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[int], List[int]]]],
+):
+    if params.method in ("attention-decoder", "rnn-lm"):
+        # Set it to False since there are too many logs.
+        enable_log = False
+    else:
+        enable_log = True
+    if test_set_name == "test-eval2000":
+        subsets = {"callhome": "en_", "swbd": "sw_", "avg": "*"}
+    elif test_set_name == "test-rt03":
+        subsets = {"fisher": "fsh_", "swbd": "sw_", "avg": "*"}
+    else:
+        raise NotImplementedError(f"No implementation for testset {test_set_name}")
+    for subset, prefix in subsets.items():
+        test_set_wers = dict()
+        for key, results in results_dict.items():
+            recog_path = params.exp_dir / f"recogs-{test_set_name}-{subset}-{key}.txt"
+            results = post_processing(results)
+            results = (
+                sorted(list(filter(lambda x: x[0].startswith(prefix), results)))
+                if subset != "avg"
+                else sorted(results)
+            )
+            store_transcripts(filename=recog_path, texts=results)
+            if enable_log:
+                logging.info(f"The transcripts are stored in {recog_path}")
+
+            # The following prints out WERs, per-word error statistics and aligned
+            # ref/hyp pairs.
+            errs_filename = params.exp_dir / f"errs-{test_set_name}-{subset}-{key}.txt"
+            with open(errs_filename, "w") as f:
+                wer = write_error_stats(
+                    f,
+                    f"{test_set_name}-{subset}-{key}",
+                    results,
+                    enable_log=enable_log,
+                    sclite_mode=True,
+                )
+                test_set_wers[key] = wer
+
+            if enable_log:
+                logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+        test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+        errs_info = params.exp_dir / f"wer-summary-{test_set_name}-{subset}.txt"
+        with open(errs_info, "w") as f:
+            print("settings\tWER", file=f)
+            for key, val in test_set_wers:
+                print("{}\t{}".format(key, val), file=f)
+
+        s = "\nFor {}-{}, WER of different settings are:\n".format(
+            test_set_name, subset
+        )
+        note = "\tbest for {}".format(test_set_name)
+        for key, val in test_set_wers:
+            s += "{}\t{}{}\n".format(key, val, note)
+            note = ""
+        logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    SwitchBoardAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+    args.lm_dir = Path(args.lm_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    graph_compiler = BpeCtcTrainingGraphCompiler(
+        params.lang_dir,
+        device=device,
+        sos_token="<sos/eos>",
+        eos_token="<sos/eos>",
+    )
+    sos_id = graph_compiler.sos_id
+    eos_id = graph_compiler.eos_id
+
+    params.num_classes = num_classes
+    params.sos_id = sos_id
+    params.eos_id = eos_id
+
+    if params.method == "ctc-decoding":
+        HLG = None
+        H = k2.ctc_topo(
+            max_token=max_token_id,
+            modified=False,
+            device=device,
+        )
+        bpe_model = spm.SentencePieceProcessor()
+        bpe_model.load(str(params.lang_dir / "bpe.model"))
+    else:
+        H = None
+        bpe_model = None
+        HLG = k2.Fsa.from_dict(
+            torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
+        )
+        assert HLG.requires_grad is False
+
+        if not hasattr(HLG, "lm_scores"):
+            HLG.lm_scores = HLG.scores.clone()
+
+    if params.method in (
+        "nbest-rescoring",
+        "whole-lattice-rescoring",
+        "attention-decoder",
+        "rnn-lm",
+    ):
+        if not (params.lm_dir / "G_4_gram.pt").is_file():
+            logging.info("Loading G_4_gram.fst.txt")
+            logging.warning("It may take 8 minutes.")
+            with open(params.lm_dir / "G_4_gram.fst.txt") as f:
+                first_word_disambig_id = lexicon.word_table["#0"]
+
+                G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+                # G.aux_labels is not needed in later computations, so
+                # remove it here.
+                del G.aux_labels
+                # CAUTION: The following line is crucial.
+                # Arcs entering the back-off state have label equal to #0.
+                # We have to change it to 0 here.
+                G.labels[G.labels >= first_word_disambig_id] = 0
+                # See https://github.com/k2-fsa/k2/issues/874
+                # for why we need to set G.properties to None
+                G.__dict__["_properties"] = None
+                G = k2.Fsa.from_fsas([G]).to(device)
+                G = k2.arc_sort(G)
+                # Save a dummy value so that it can be loaded in C++.
+                # See https://github.com/pytorch/pytorch/issues/67902
+                # for why we need to do this.
+                G.dummy = 1
+
+                torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
+        else:
+            logging.info("Loading pre-compiled G_4_gram.pt")
+            d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
+            G = k2.Fsa.from_dict(d)
+
+        if params.method in [
+            "whole-lattice-rescoring",
+            "attention-decoder",
+            "rnn-lm",
+        ]:
+            # Add epsilon self-loops to G as we will compose
+            # it with the whole lattice later
+            G = k2.add_epsilon_self_loops(G)
+            G = k2.arc_sort(G)
+            G = G.to(device)
+
+        # G.lm_scores is used to replace HLG.lm_scores during
+        # LM rescoring.
+        G.lm_scores = G.scores.clone()
+    else:
+        G = None
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=params.vgg_frontend,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        model = load_averaged_model(
+            params.exp_dir, model, params.epoch, params.avg, device
+        )
+
+    model.to(device)
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    rnn_lm_model = None
+    if params.method == "rnn-lm":
+        rnn_lm_model = RnnLmModel(
+            vocab_size=params.num_classes,
+            embedding_dim=params.rnn_lm_embedding_dim,
+            hidden_dim=params.rnn_lm_hidden_dim,
+            num_layers=params.rnn_lm_num_layers,
+            tie_weights=params.rnn_lm_tie_weights,
+        )
+        if params.rnn_lm_avg == 1:
+            load_checkpoint(
+                f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt",
+                rnn_lm_model,
+            )
+            rnn_lm_model.to(device)
+        else:
+            rnn_lm_model = load_averaged_model(
+                params.rnn_lm_exp_dir,
+                rnn_lm_model,
+                params.rnn_lm_epoch,
+                params.rnn_lm_avg,
+                device,
+            )
+        rnn_lm_model.eval()
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    switchboard = SwitchBoardAsrDataModule(args)
+
+    test_eval2000_cuts = switchboard.test_eval2000_cuts().trim_to_supervisions(
+        keep_all_channels=True
+    )
+    # test_rt03_cuts = switchboard.test_rt03_cuts().trim_to_supervisions(
+    #     keep_all_channels=True
+    # )
+
+    test_eval2000_dl = switchboard.test_dataloaders(test_eval2000_cuts)
+    # test_rt03_dl = switchboard.test_dataloaders(test_rt03_cuts)
+
+    # test_sets = ["test-eval2000", "test-rt03"]
+    # test_dl = [test_eval2000_dl, test_rt03_dl]
+    test_sets = ["test-eval2000"]
+    test_dl = [test_eval2000_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            rnn_lm_model=rnn_lm_model,
+            HLG=HLG,
+            H=H,
+            bpe_model=bpe_model,
+            word_table=lexicon.word_table,
+            G=G,
+            sos_id=sos_id,
+            eos_id=eos_id,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/conformer_ctc/export.py b/egs/swbd/ASR/conformer_ctc/export.py
new file mode 100755
index 000000000..1bb6277ad
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/export.py
@@ -0,0 +1,163 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from conformer import Conformer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=98,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=55,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""It contains language related input files such as "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=True,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_feat_batchnorm": True,
+            "attention_dim": 512,
+            "nhead": 8,
+            "num_decoder_layers": 6,
+        }
+    )
+    return params
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    logging.info(params)
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.load_state_dict(average_checkpoints(filenames))
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/swbd/ASR/conformer_ctc/label_smoothing.py b/egs/swbd/ASR/conformer_ctc/label_smoothing.py
new file mode 120000
index 000000000..e9d239fff
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/label_smoothing.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/label_smoothing.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/conformer_ctc/pretrained.py b/egs/swbd/ASR/conformer_ctc/pretrained.py
new file mode 120000
index 000000000..526bc9678
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/pretrained.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/pretrained.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/conformer_ctc/sclite_scoring.py b/egs/swbd/ASR/conformer_ctc/sclite_scoring.py
new file mode 100755
index 000000000..0383c4d71
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/sclite_scoring.py
@@ -0,0 +1,148 @@
+#!/usr/bin/env python3
+# Copyright 2021 Jiayu Du
+# Copyright 2022 Johns Hopkins University (Author: Guanbo Wang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import os
+
+conversational_filler = [
+    "UH",
+    "UHH",
+    "UM",
+    "EH",
+    "MM",
+    "HM",
+    "AH",
+    "HUH",
+    "HA",
+    "ER",
+    "OOF",
+    "HEE",
+    "ACH",
+    "EEE",
+    "EW",
+    "MHM",
+    "HUM",
+    "AW",
+    "OH",
+    "HMM",
+    "UMM",
+]
+unk_tags = ["<UNK>", "<unk>"]
+switchboard_garbage_utterance_tags = [
+    "[LAUGHTER]",
+    "[NOISE]",
+    "[VOCALIZED-NOISE]",
+    "[SILENCE]",
+]
+non_scoring_words = (
+    conversational_filler + unk_tags + switchboard_garbage_utterance_tags
+)
+
+
+def asr_text_post_processing(text: str) -> str:
+    # 1. convert to uppercase
+    text = text.upper()
+
+    # 2. remove non-scoring words from evaluation
+    remaining_words = []
+    text_split = text.split()
+    word_to_skip = 0
+    for idx, word in enumerate(text_split):
+        if word_to_skip > 0:
+            word_to_skip -= 1
+            continue
+        if word in non_scoring_words:
+            continue
+        elif word == "CANCELLED":
+            remaining_words.append("CANCELED")
+            continue
+        elif word == "AIRFLOW":
+            remaining_words.append("AIR")
+            remaining_words.append("FLOW")
+            continue
+        elif word == "PHD":
+            remaining_words.append("P")
+            remaining_words.append("H")
+            remaining_words.append("D")
+            continue
+        elif word == "UCLA":
+            remaining_words.append("U")
+            remaining_words.append("C")
+            remaining_words.append("L")
+            remaining_words.append("A")
+            continue
+        elif word == "ONTO":
+            remaining_words.append("ON")
+            remaining_words.append("TO")
+            continue
+        elif word == "DAY":
+            try:
+                if text_split[idx + 1] == "CARE":
+                    remaining_words.append("DAYCARE")
+                word_to_skip = 1
+            except:
+                remaining_words.append(word)
+            continue
+        remaining_words.append(word)
+
+    return " ".join(remaining_words)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="This script evaluates GigaSpeech ASR result via"
+        "SCTK's tool sclite"
+    )
+    parser.add_argument(
+        "ref",
+        type=str,
+        help="sclite's standard transcription(trn) reference file",
+    )
+    parser.add_argument(
+        "hyp",
+        type=str,
+        help="sclite's standard transcription(trn) hypothesis file",
+    )
+    parser.add_argument(
+        "work_dir",
+        type=str,
+        help="working dir",
+    )
+    args = parser.parse_args()
+
+    if not os.path.isdir(args.work_dir):
+        os.mkdir(args.work_dir)
+
+    REF = os.path.join(args.work_dir, "REF")
+    HYP = os.path.join(args.work_dir, "HYP")
+    RESULT = os.path.join(args.work_dir, "RESULT")
+
+    for io in [(args.ref, REF), (args.hyp, HYP)]:
+        with open(io[0], "r", encoding="utf8") as fi:
+            with open(io[1], "w+", encoding="utf8") as fo:
+                for line in fi:
+                    line = line.strip()
+                    if line:
+                        cols = line.split()
+                        text = asr_text_post_processing(" ".join(cols[0:-1]))
+                        uttid_field = cols[-1]
+                        print(f"{text} {uttid_field}", file=fo)
+
+    # GigaSpeech's uttid comforms to swb
+    os.system(f"sclite -r {REF} trn -h {HYP} trn -i swb | tee {RESULT}")
diff --git a/egs/swbd/ASR/conformer_ctc/subsampling.py b/egs/swbd/ASR/conformer_ctc/subsampling.py
new file mode 120000
index 000000000..16354dc73
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/subsampling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/subsampling.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/conformer_ctc/test_label_smoothing.py b/egs/swbd/ASR/conformer_ctc/test_label_smoothing.py
new file mode 100755
index 000000000..5d4438fd1
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/test_label_smoothing.py
@@ -0,0 +1,52 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from distutils.version import LooseVersion
+
+import torch
+from label_smoothing import LabelSmoothingLoss
+
+torch_ver = LooseVersion(torch.__version__)
+
+
+def test_with_torch_label_smoothing_loss():
+    if torch_ver < LooseVersion("1.10.0"):
+        print(f"Current torch version: {torch_ver}")
+        print("Please use torch >= 1.10 to run this test - skipping")
+        return
+    torch.manual_seed(20211105)
+    x = torch.rand(20, 30, 5000)
+    tgt = torch.randint(low=-1, high=x.size(-1), size=x.shape[:2])
+    for reduction in ["none", "sum", "mean"]:
+        custom_loss_func = LabelSmoothingLoss(
+            ignore_index=-1, label_smoothing=0.1, reduction=reduction
+        )
+        custom_loss = custom_loss_func(x, tgt)
+
+        torch_loss_func = torch.nn.CrossEntropyLoss(
+            ignore_index=-1, reduction=reduction, label_smoothing=0.1
+        )
+        torch_loss = torch_loss_func(x.reshape(-1, x.size(-1)), tgt.reshape(-1))
+        assert torch.allclose(custom_loss, torch_loss)
+
+
+def main():
+    test_with_torch_label_smoothing_loss()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/conformer_ctc/test_subsampling.py b/egs/swbd/ASR/conformer_ctc/test_subsampling.py
new file mode 100755
index 000000000..81fa234dd
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/test_subsampling.py
@@ -0,0 +1,48 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+from subsampling import Conv2dSubsampling, VggSubsampling
+
+
+def test_conv2d_subsampling():
+    N = 3
+    odim = 2
+
+    for T in range(7, 19):
+        for idim in range(7, 20):
+            model = Conv2dSubsampling(idim=idim, odim=odim)
+            x = torch.empty(N, T, idim)
+            y = model(x)
+            assert y.shape[0] == N
+            assert y.shape[1] == ((T - 1) // 2 - 1) // 2
+            assert y.shape[2] == odim
+
+
+def test_vgg_subsampling():
+    N = 3
+    odim = 2
+
+    for T in range(7, 19):
+        for idim in range(7, 20):
+            model = VggSubsampling(idim=idim, odim=odim)
+            x = torch.empty(N, T, idim)
+            y = model(x)
+            assert y.shape[0] == N
+            assert y.shape[1] == ((T - 1) // 2 - 1) // 2
+            assert y.shape[2] == odim
diff --git a/egs/swbd/ASR/conformer_ctc/test_transformer.py b/egs/swbd/ASR/conformer_ctc/test_transformer.py
new file mode 120000
index 000000000..8b0990ec6
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/test_transformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/test_transformer.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/conformer_ctc/train.py b/egs/swbd/ASR/conformer_ctc/train.py
new file mode 100755
index 000000000..7f1eebbcf
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/train.py
@@ -0,0 +1,814 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+# Modified by Zengrui Jin for the SwitchBoard corpus
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./conformer_ctc/train.py \
+     --exp-dir ./conformer_ctc/exp \
+     --world-size 4 \
+     --max-duration 200 \
+     --num-epochs 20
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import SwitchBoardAsrDataModule
+from conformer import Conformer
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.graph_compiler import CtcTrainingGraphCompiler
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    encode_supervisions,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=98,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        conformer_ctc/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="conformer_ctc/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_bpe_500",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--att-rate",
+        type=float,
+        default=0.8,
+        help="""The attention rate.
+        The total loss is (1 -  att_rate) * ctc_loss + att_rate * att_loss
+        """,
+    )
+
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=6,
+        help="""Number of decoder layer of transformer decoder.
+        Setting this to 0 will not create the decoder at all (pure CTC model)
+        """,
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - use_feat_batchnorm: Normalization for the input features, can be a
+                              boolean indicating whether to do batch
+                              normalization, or a float which means just scaling
+                              the input features with this float value.
+                              If given a float value, we will remove batchnorm
+                              layer in `ConvolutionModule` as well.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - head: Number of heads of multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - beam_size: It is used in k2.ctc_loss
+
+        - reduction: It is used in k2.ctc_loss
+
+        - use_double_scores: It is used in k2.ctc_loss
+
+        - weight_decay:  The weight_decay for the optimizer.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "use_feat_batchnorm": True,
+            "attention_dim": 512,
+            "nhead": 8,
+            # parameters for loss
+            "beam_size": 10,
+            "reduction": "sum",
+            "use_double_scores": True,
+            # parameters for Noam
+            "weight_decay": 1e-6,
+            "warm_step": 80000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    batch: dict,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      graph_compiler:
+        It is used to build a decoding graph from a ctc topo and training
+        transcript. The training transcript is contained in the given `batch`,
+        while the ctc topo is built when this compiler is instantiated.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = graph_compiler.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    with torch.set_grad_enabled(is_training):
+        nnet_output, encoder_memory, memory_mask = model(feature, supervisions)
+        # nnet_output is (N, T, C)
+
+    # NOTE: We need `encode_supervisions` to sort sequences with
+    # different duration in decreasing order, required by
+    # `k2.intersect_dense` called in `k2.ctc_loss`
+    supervision_segments, texts = encode_supervisions(
+        supervisions, subsampling_factor=params.subsampling_factor
+    )
+
+    if isinstance(graph_compiler, BpeCtcTrainingGraphCompiler):
+        # Works with a BPE model
+        token_ids = graph_compiler.texts_to_ids(texts)
+        decoding_graph = graph_compiler.compile(token_ids)
+    elif isinstance(graph_compiler, CtcTrainingGraphCompiler):
+        # Works with a phone lexicon
+        decoding_graph = graph_compiler.compile(texts)
+    else:
+        raise ValueError(f"Unsupported type of graph compiler: {type(graph_compiler)}")
+
+    dense_fsa_vec = k2.DenseFsaVec(
+        nnet_output,
+        supervision_segments,
+        allow_truncate=params.subsampling_factor - 1,
+    )
+
+    ctc_loss = k2.ctc_loss(
+        decoding_graph=decoding_graph,
+        dense_fsa_vec=dense_fsa_vec,
+        output_beam=params.beam_size,
+        reduction=params.reduction,
+        use_double_scores=params.use_double_scores,
+    )
+
+    if params.att_rate != 0.0:
+        with torch.set_grad_enabled(is_training):
+            mmodel = model.module if hasattr(model, "module") else model
+            # Note: We need to generate an unsorted version of token_ids
+            # `encode_supervisions()` called above sorts text, but
+            # encoder_memory and memory_mask are not sorted, so we
+            # use an unsorted version `supervisions["text"]` to regenerate
+            # the token_ids
+            #
+            # See https://github.com/k2-fsa/icefall/issues/97
+            # for more details
+            unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"])
+            att_loss = mmodel.decoder_forward(
+                encoder_memory,
+                memory_mask,
+                token_ids=unsorted_token_ids,
+                sos_id=graph_compiler.sos_id,
+                eos_id=graph_compiler.eos_id,
+            )
+        loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss
+    else:
+        loss = ctc_loss
+        att_loss = torch.tensor([0])
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = supervision_segments[:, 2].sum().item()
+    info["ctc_loss"] = ctc_loss.detach().cpu().item()
+    if params.att_rate != 0.0:
+        info["att_loss"] = att_loss.detach().cpu().item()
+
+    info["loss"] = loss.detach().cpu().item()
+
+    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
+    info["utterances"] = feature.size(0)
+    # averaged input duration in frames over utterances
+    info["utt_duration"] = supervisions["num_frames"].sum().item()
+    # averaged padding proportion over utterances
+    info["utt_pad_proportion"] = (
+        ((feature.size(1) - supervisions["num_frames"]) / feature.size(1)).sum().item()
+    )
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      graph_compiler:
+        It is used to convert transcripts to FSAs.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            batch=batch,
+            graph_compiler=graph_compiler,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    lexicon = Lexicon(params.lang_dir)
+    max_token_id = max(lexicon.tokens)
+    num_classes = max_token_id + 1  # +1 for the blank
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    if "lang_bpe" in str(params.lang_dir):
+        graph_compiler = BpeCtcTrainingGraphCompiler(
+            params.lang_dir,
+            device=device,
+            sos_token="<sos/eos>",
+            eos_token="<sos/eos>",
+        )
+    elif "lang_phone" in str(params.lang_dir):
+        assert params.att_rate == 0, (
+            "Attention decoder training does not support phone lang dirs "
+            "at this time due to a missing <sos/eos> symbol. Set --att-rate=0 "
+            "for pure CTC training when using a phone-based lang dir."
+        )
+        assert params.num_decoder_layers == 0, (
+            "Attention decoder training does not support phone lang dirs "
+            "at this time due to a missing <sos/eos> symbol. "
+            "Set --num-decoder-layers=0 for pure CTC training when using "
+            "a phone-based lang dir."
+        )
+        graph_compiler = CtcTrainingGraphCompiler(
+            lexicon,
+            device=device,
+        )
+        # Manually add the sos/eos ID with their default values
+        # from the BPE recipe which we're adapting here.
+        graph_compiler.sos_id = 1
+        graph_compiler.eos_id = 1
+    else:
+        raise ValueError(
+            f"Unsupported type of lang dir (we expected it to have "
+            f"'lang_bpe' or 'lang_phone' in its name): {params.lang_dir}"
+        )
+
+    logging.info("About to create model")
+    model = Conformer(
+        num_features=params.feature_dim,
+        nhead=params.nhead,
+        d_model=params.attention_dim,
+        num_classes=num_classes,
+        subsampling_factor=params.subsampling_factor,
+        num_decoder_layers=params.num_decoder_layers,
+        vgg_frontend=False,
+        use_feat_batchnorm=params.use_feat_batchnorm,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        model = DDP(model, device_ids=[rank])
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+        weight_decay=params.weight_decay,
+    )
+
+    if checkpoints:
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    switchboard = SwitchBoardAsrDataModule(args)
+
+    train_cuts = switchboard.train_all_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = switchboard.train_dataloaders(train_cuts)
+
+    valid_cuts = switchboard.dev_cuts()
+    valid_dl = switchboard.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        graph_compiler=graph_compiler,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: BpeCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                batch=batch,
+                graph_compiler=graph_compiler,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    SwitchBoardAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/conformer_ctc/transformer.py b/egs/swbd/ASR/conformer_ctc/transformer.py
new file mode 120000
index 000000000..1c3f43fcf
--- /dev/null
+++ b/egs/swbd/ASR/conformer_ctc/transformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/transformer.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/local/compile_hlg.py b/egs/swbd/ASR/local/compile_hlg.py
new file mode 120000
index 000000000..471aa7fb4
--- /dev/null
+++ b/egs/swbd/ASR/local/compile_hlg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_hlg.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/local/compile_lg.py b/egs/swbd/ASR/local/compile_lg.py
new file mode 120000
index 000000000..462d6d3fb
--- /dev/null
+++ b/egs/swbd/ASR/local/compile_lg.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_lg.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/local/compute_fbank_eval2000.py b/egs/swbd/ASR/local/compute_fbank_eval2000.py
new file mode 100755
index 000000000..d446e8ff3
--- /dev/null
+++ b/egs/swbd/ASR/local/compute_fbank_eval2000.py
@@ -0,0 +1,139 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# Modified     2023  The Chinese University of Hong Kong (author: Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the SwitchBoard dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import sentencepiece as spm
+import torch
+from filter_cuts import filter_cuts
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to the bpe.model. If not None, we will remove short and
+        long utterances before extracting features""",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="""Perturb speed with factor 0.9 and 1.1 on train subset.""",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_switchboard(
+    dir_name: str,
+    bpe_model: Optional[str] = None,
+    dataset: Optional[str] = None,
+    perturb_speed: Optional[bool] = True,
+):
+    src_dir = Path(f"data/manifests/{dir_name}")
+    output_dir = Path(f"data/fbank/{dir_name}")
+    num_jobs = min(1, os.cpu_count())
+    num_mel_bins = 80
+
+    if bpe_model:
+        logging.info(f"Loading {bpe_model}")
+        sp = spm.SentencePieceProcessor()
+        sp.load(bpe_model)
+
+    if dataset is None:
+        dataset_parts = ("all",)
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    prefix = dir_name
+    suffix = "jsonl.gz"
+    manifests = {
+        "eval2000": "data/manifests/eval2000/eval2000_cuts_all_trimmed.jsonl.gz",
+    }
+    assert manifests is not None
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=16000))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        partition = "all"
+        cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+        print(cuts_filename)
+        if (output_dir / cuts_filename).is_file():
+            logging.info(f"{prefix} already exists - skipping.")
+            return
+        logging.info(f"Processing {prefix}")
+        cut_set = CutSet.from_file(manifests[prefix]).resample(16000)
+
+        cut_set = cut_set.compute_and_store_features(
+            extractor=extractor,
+            storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+            # when an executor is specified, make more partitions
+            num_jobs=num_jobs if ex is None else 80,
+            executor=ex,
+            storage_type=LilcomChunkyWriter,
+        )
+        cut_set = cut_set.trim_to_supervisions(keep_overlapping=False)
+        cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    compute_fbank_switchboard(
+        dir_name="eval2000",
+        bpe_model=args.bpe_model,
+        dataset=args.dataset,
+        perturb_speed=args.perturb_speed,
+    )
diff --git a/egs/swbd/ASR/local/compute_fbank_swbd.py b/egs/swbd/ASR/local/compute_fbank_swbd.py
new file mode 100755
index 000000000..dd82220c0
--- /dev/null
+++ b/egs/swbd/ASR/local/compute_fbank_swbd.py
@@ -0,0 +1,163 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# Modified     2023  The Chinese University of Hong Kong (author: Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+"""
+This file computes fbank features of the SwitchBoard dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import sentencepiece as spm
+import torch
+from filter_cuts import filter_cuts
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor, str2bool
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to the bpe.model. If not None, we will remove short and
+        long utterances before extracting features""",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="""Perturb speed with factor 0.9 and 1.1 on train subset.""",
+    )
+
+    parser.add_argument(
+        "--split-index",
+        type=int,
+        required=True,
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_switchboard(
+    dir_name: str,
+    split_index: int,
+    bpe_model: Optional[str] = None,
+    dataset: Optional[str] = None,
+    perturb_speed: Optional[bool] = True,
+):
+    src_dir = Path(f"data/manifests/{dir_name}")
+    output_dir = Path(f"data/fbank/{dir_name}_split16")
+    num_jobs = min(1, os.cpu_count())
+    num_mel_bins = 80
+
+    if bpe_model:
+        logging.info(f"Loading {bpe_model}")
+        sp = spm.SentencePieceProcessor()
+        sp.load(bpe_model)
+
+    if dataset is None:
+        dataset_parts = ("all",)
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    prefix = dir_name
+    suffix = "jsonl.gz"
+    split_dir = Path("data/manifests/swbd_split16/")
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=16000))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        partition = "all"
+        cuts_filename = (
+            f"{prefix}_cuts_{partition}.{str(split_index).zfill(2)}.{suffix}"
+        )
+        print(cuts_filename)
+        if (output_dir / cuts_filename).is_file():
+            logging.info(f"{prefix} already exists - skipping.")
+            return
+        logging.info(f"Processing {prefix}")
+        cut_set = (
+            CutSet.from_file(
+                split_dir
+                / f"swbd_train_all_trimmed.{str(split_index).zfill(2)}.jsonl.gz"
+            )
+            .resample(16000)
+            .to_eager()
+            .filter(lambda c: c.duration > 2.0)
+        )
+
+        if bpe_model:
+            cut_set = filter_cuts(cut_set, sp)
+        if perturb_speed:
+            logging.info(f"Doing speed perturb")
+            cut_set = cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+        cut_set = cut_set.compute_and_store_features(
+            extractor=extractor,
+            storage_path=f"{output_dir}/{prefix}_feats_{partition}_{str(split_index).zfill(2)}",
+            # when an executor is specified, make more partitions
+            num_jobs=num_jobs if ex is None else 80,
+            executor=ex,
+            storage_type=LilcomChunkyWriter,
+        )
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False,
+            min_duration=None,
+        )
+        cut_set.to_file(output_dir / cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+
+    compute_fbank_switchboard(
+        dir_name="swbd",
+        split_index=args.split_index,
+        bpe_model=args.bpe_model,
+        dataset=args.dataset,
+        perturb_speed=args.perturb_speed,
+    )
diff --git a/egs/swbd/ASR/local/convert_transcript_words_to_tokens.py b/egs/swbd/ASR/local/convert_transcript_words_to_tokens.py
new file mode 100755
index 000000000..a8d5117c9
--- /dev/null
+++ b/egs/swbd/ASR/local/convert_transcript_words_to_tokens.py
@@ -0,0 +1,103 @@
+#!/usr/bin/env python3
+
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+"""
+Convert a transcript file containing words to a corpus file containing tokens
+for LM training with the help of a lexicon.
+
+If the lexicon contains phones, the resulting LM will be a phone LM; If the
+lexicon contains word pieces, the resulting LM will be a word piece LM.
+
+If a word has multiple pronunciations, the one that appears first in the lexicon
+is kept; others are removed.
+
+If the input transcript is:
+
+    hello zoo world hello
+    world zoo
+    foo zoo world hellO
+
+and if the lexicon is
+
+    <UNK> SPN
+    hello h e l l o 2
+    hello h e l l o
+    world w o r l d
+    zoo z o o
+
+Then the output is
+
+    h e l l o 2 z o o w o r l d h e l l o 2
+    w o r l d z o o
+    SPN z o o w o r l d SPN
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List
+
+from generate_unique_lexicon import filter_multiple_pronunications
+
+from icefall.lexicon import read_lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="The input transcript file."
+        "We assume that the transcript file consists of "
+        "lines. Each line consists of space separated words.",
+    )
+    parser.add_argument("--lexicon", type=str, help="The input lexicon file.")
+    parser.add_argument("--oov", type=str, default="<UNK>", help="The OOV word.")
+
+    return parser.parse_args()
+
+
+def process_line(lexicon: Dict[str, List[str]], line: str, oov_token: str) -> None:
+    """
+    Args:
+      lexicon:
+        A dict containing pronunciations. Its keys are words and values
+        are pronunciations (i.e., tokens).
+      line:
+        A line of transcript consisting of space(s) separated words.
+      oov_token:
+        The pronunciation of the oov word if a word in `line` is not present
+        in the lexicon.
+    Returns:
+      Return None.
+    """
+    s = ""
+    words = line.strip().split()
+    for i, w in enumerate(words):
+        tokens = lexicon.get(w, oov_token)
+        s += " ".join(tokens)
+        s += " "
+    print(s.strip())
+
+
+def main():
+    args = get_args()
+    assert Path(args.lexicon).is_file()
+    assert Path(args.transcript).is_file()
+    assert len(args.oov) > 0
+
+    # Only the first pronunciation of a word is kept
+    lexicon = filter_multiple_pronunications(read_lexicon(args.lexicon))
+
+    lexicon = dict(lexicon)
+
+    assert args.oov in lexicon
+
+    oov_token = lexicon[args.oov]
+
+    with open(args.transcript) as f:
+        for line in f:
+            process_line(lexicon=lexicon, line=line, oov_token=oov_token)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/local/dict.patch b/egs/swbd/ASR/local/dict.patch
new file mode 100644
index 000000000..12c63d612
--- /dev/null
+++ b/egs/swbd/ASR/local/dict.patch
@@ -0,0 +1,380 @@
+1d0
+<  file: $SWB/data/dictionary/sw-ms98-dict.text
+8645a8646
+> uh-hum ah m hh ah m
+9006c9007
+< April ey p r ih l
+---
+> April ey p r ax l
+9144d9144
+< B ay zh aa n iy z
+9261c9261
+< Battle b ae t el
+---
+> Battle b ae t ax l
+10014a10015
+> Chevy sh eh v iy
+10211a10213
+> Colorado k ao l ax r aa d ow
+10212a10215
+> Colorado' k ao l ax r aa d ow z
+10370c10373
+< Creek k r ih k
+---
+> Creek k r iy k
+10889a10893
+> Eleven ax l eh v ih n
+10951c10955
+< Erie ih r iy
+---
+> Erie iy r iy
+11183c11187
+< Forever f ax r eh v er
+---
+> Forever f er eh v er
+11231a11236
+> Friday f r ay d iy
+11744a11750
+> History hh ih s t r iy
+12004a12011,12012
+> Israel ih z r ih l
+> Israel's ih z r ih l z
+12573a12582
+> Lincoln l ih ng k ih n
+12574a12584
+> Lincolns l ih ng k ih n z
+13268c13278
+< NAACP eh ey ey s iy p iy
+---
+> NAACP eh n ey ey s iy p iy
+13286c13296
+< NIT eh ay t iy
+---
+> NIT eh n ay t iy
+13292c13302
+< NTSC eh t iy eh s s iy
+---
+> NTSC eh n t iy eh s s iy
+14058a14069
+> Quarter k ow r t er
+14059a14071
+> Quarterback k ow r t er b ae k
+14060a14073
+> Quarters k ow r t er z
+14569a14583
+> Science s ay n s
+15087a15102
+> Sunday s ah n d iy
+15088a15104
+> Sunday's s ah n d iy z
+15089a15106
+> Sundays s ah n d iy z
+15290,15291c15307,15308
+< Texan t eh k sh ih n
+< Texan's t eh k sh ih n s
+---
+> Texan t eh k s ih n
+> Texan's t eh k s ih n s
+15335a15353
+> Thousands th aw z ih n z
+15739c15757
+< Waco w ae k ow
+---
+> Waco w ey k ow
+15841a15860
+> Weekends w iy k eh n z
+16782a16802
+> acceptable eh k s eh p ax b ax l
+16833a16854
+> accounting ax k aw n ih ng
+16948a16970
+> address ax d r eh s
+17281a17304
+> already aa r d iy
+17315a17339
+> am m
+17709a17734
+> asked ae s t
+17847a17873
+> attorney ih t er n iy
+17919a17946
+> autopilot ao t ow p ay l ih t
+17960a17988
+> awfully ao f l iy
+18221a18250
+> basketball b ae s k ax b ao l
+18222a18252
+> basketball's b ae s k ax b ao l z
+18302a18333
+> become b ah k ah m
+18303a18335
+> becomes b iy k ah m z
+18344a18377
+> began b ax g en n
+18817c18850
+< bottle b aa t el
+---
+> bottle b aa t ax l
+19332,19333c19365,19367
+< camera's k ae m ax r ax z
+< cameras k ae m ax r ax z
+---
+> camera k ae m r ax
+> camera's k ae m r ax z
+> cameras k ae m r ax z
+19411a19446
+> capital k ae p ax l
+19505a19541
+> carrying k ae r ih ng
+20316a20353,20354
+> combination k aa m ih n ey sh ih n
+> combinations k aa m ih n ey sh ih n z
+20831a20870
+> contracts k aa n t r ae k s
+21010a21050
+> costs k ao s
+21062a21103
+> county k aw n iy
+21371a21413
+> cultural k ao l ch ax r ax l
+21372a21415
+> culturally k ao l ch ax r ax l iy
+21373a21417
+> culture k ao l ch er
+21375a21420
+> cultures k ao l ch er z
+21543a21589
+> data d ey t ax
+22097a22144
+> differently d ih f ax r ih n t l iy
+22972a23020
+> effects ax f eh k t s
+23016a23065
+> election ax l eh k sh ih n
+23018a23068
+> elections ax l eh k sh ih n z
+23052a23103
+> eleven ax l eh v ih n
+23242a23294
+> enjoyable ae n jh oy ax b ax l
+23248a23301
+> enjoys ae n jh oy z
+23293a23347
+> entire ih n t ay r
+23295a23350,23351
+> entirely ih n t ay r l iy
+> entirety ih n t ay r t iy
+23745a23802
+> extra eh k s t er
+23818a23876
+> facts f ae k s
+24508c24566
+< forever f ax r eh v er
+---
+> forever f er eh v er
+24514c24572
+< forget f ow r g eh t
+---
+> forget f er r g eh t
+24521a24580
+> forgot f er r g aa t
+24522a24582
+> forgotten f er r g aa t ax n
+24563a24624
+> forward f ow er d
+24680a24742
+> frightening f r ay t n ih ng
+24742a24805
+> full-time f ax l t ay m
+24862a24926
+> garage g r aa jh
+25218a25283
+> grandmother g r ae m ah dh er
+25790a25856
+> heavily hh eh v ax l iy
+25949a26016
+> history hh ih s t r iy
+26038a26106
+> honestly aa n ax s t l iy
+26039a26108
+> honesty aa n ax s t iy
+26099a26169
+> horror hh ow r
+26155a26226
+> houses hh aw z ih z
+26184c26255
+< huh-uh hh ah hh ah
+---
+> huh-uh ah hh ah
+26189c26260
+< hum-um hh m hh m
+---
+> hum-um ah m hh ah m
+26236a26308
+> hunting hh ah n ih ng
+26307a26380,26381
+> ideal ay d iy l
+> idealist ay d iy l ih s t
+26369a26444
+> imagine m ae jh ih n
+26628a26704
+> individuals ih n d ih v ih jh ax l z
+26968a27045
+> interest ih n t r ih s t
+27184a27262
+> it'd ih d
+27702a27781
+> lead l iy d
+28378a28458
+> mandatory m ae n d ih t ow r iy
+28885a28966
+> minute m ih n ih t
+29167a29249
+> mountains m aw t n z
+29317a29400
+> mysteries m ih s t r iy z
+29318a29402
+> mystery m ih s t r iy
+29470a29555
+> nervous n er v ih s
+29578,29580c29663,29665
+< nobody n ow b aa d iy
+< nobody'll n ow b aa d iy l
+< nobody's n ow b aa d iy z
+---
+> nobody n ow b ah d iy
+> nobody'll n ow b ah d iy l
+> nobody's n ow b ah d iy z
+29712a29798
+> nuclear n uw k l iy r
+29938a30025
+> onto aa n t ax
+30051a30139
+> originally ax r ih jh ax l iy
+30507a30596
+> particularly p er t ih k y ax l iy
+30755a30845
+> perfectly p er f ih k l iy
+30820a30911
+> personally p er s n ax l iy
+30915a31007
+> physically f ih z ih k l iy
+30986a31079
+> pilot p ay l ih t
+30987a31081
+> pilot's p ay l ih t s
+31227a31322
+> police p l iy s
+31513a31609
+> prefer p er f er
+31553a31650
+> prepare p r ax p ey r
+31578a31676
+> prescription p er s k r ih p sh ih n
+31579a31678
+> prescriptions p er s k r ih p sh ih n z
+31770a31870
+> products p r aa d ax k s
+31821a31922
+> projects p r aa jh eh k s
+31908a32010
+> protect p er t eh k t
+31909a32012
+> protected p er t eh k t ih d
+31911a32015
+> protection p er t eh k sh ih n
+31914a32019
+> protection p er t eh k t ih v
+32149a32255
+> quarter k ow r t er
+32414a32521
+> read r iy d
+32785a32893
+> rehabilitation r iy ax b ih l ih t ey sh ih n
+33150a33259
+> resource r ih s ow r s
+33151a33261
+> resources r iy s ow r s ih z
+33539c33649
+< roots r uh t s
+---
+> roots r uw t s
+33929a34040
+> science s ay n s
+34315a34427
+> seventy s eh v ih n iy
+34319,34320c34431,34432
+< severe s ax v iy r
+< severely s ax v iy r l iy
+---
+> severe s ih v iy r
+> severely s ih v iy r l iy
+35060a35173
+> software s ao f w ey r
+35083a35197
+> solid s ao l ih d
+35084a35199
+> solidly s ao l ih d l iy
+35750a35866
+> stood s t ih d
+35854a35971
+> strictly s t r ih k l iy
+35889c36006
+< stronger s t r ao ng er
+---
+> stronger s t r ao ng g er
+36192a36310,36311
+> supposed s p ow z
+> supposed s p ow s
+36510a36630
+> tastes t ey s
+36856a36977
+> thoroughly th er r l iy
+36866a36988
+> thousands th aw z ih n z
+37081c37203
+< toots t uh t s
+---
+> toots t uw t s
+37157a37280
+> toward t w ow r d
+37158a37282
+> towards t w ow r d z
+37564a37689
+> twenties t w eh n iy z
+37565a37691
+> twentieth t w eh n iy ih th
+37637a37764
+> unacceptable ah n ae k s eh p ax b ax l
+37728a37856
+> understand ah n d er s t ae n
+37860a37989
+> unless ih n l eh s
+38040a38170
+> use y uw z
+38049a38180
+> uses y uw z ih z
+38125a38257
+> various v ah r iy ih s
+38202a38335
+> versus v er s ih z
+38381c38514
+< wacko w ae k ow
+---
+> wacko w ey k ow
+38455c38588
+< wanna w aa n ax
+---
+> wanna w ah n ax
+38675c38808
+< whatnot w ah t n aa t
+---
+> whatnot w aa t n aa t
+38676a38810
+> whatsoever w aa t s ow eh v er
+38890c39024
+< wok w aa k
+---
+> wok w ao k
+38910a39045
+> wondering w ah n d r ih ng
diff --git a/egs/swbd/ASR/local/display_manifest_statistics.py b/egs/swbd/ASR/local/display_manifest_statistics.py
new file mode 100755
index 000000000..9aa204863
--- /dev/null
+++ b/egs/swbd/ASR/local/display_manifest_statistics.py
@@ -0,0 +1,125 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+See the function `remove_short_and_long_utt()` in transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest_lazy
+
+
+def main():
+    #  path = "./data/fbank/swbd_cuts_rt03.jsonl.gz"
+    path = "./data/fbank/eval2000/eval2000_cuts_all.jsonl.gz"
+    # path = "./data/fbank/swbd_cuts_all.jsonl.gz"
+
+    cuts = load_manifest_lazy(path)
+    cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+Training Cut statistics:
+╒═══════════════════════════╤═══════════╕
+│ Cuts count:               │ 167244    │
+├───────────────────────────┼───────────┤
+│ Total duration (hh:mm:ss) │ 281:01:26 │
+├───────────────────────────┼───────────┤
+│ mean                      │ 6.0       │
+├───────────────────────────┼───────────┤
+│ std                       │ 3.3       │
+├───────────────────────────┼───────────┤
+│ min                       │ 2.0       │
+├───────────────────────────┼───────────┤
+│ 25%                       │ 3.2       │
+├───────────────────────────┼───────────┤
+│ 50%                       │ 5.2       │
+├───────────────────────────┼───────────┤
+│ 75%                       │ 8.3       │
+├───────────────────────────┼───────────┤
+│ 99%                       │ 14.4      │
+├───────────────────────────┼───────────┤
+│ 99.5%                     │ 14.7      │
+├───────────────────────────┼───────────┤
+│ 99.9%                     │ 15.0      │
+├───────────────────────────┼───────────┤
+│ max                       │ 57.5      │
+├───────────────────────────┼───────────┤
+│ Recordings available:     │ 167244    │
+├───────────────────────────┼───────────┤
+│ Features available:       │ 167244    │
+├───────────────────────────┼───────────┤
+│ Supervisions available:   │ 167244    │
+╘═══════════════════════════╧═══════════╛
+Speech duration statistics:
+╒══════════════════════════════╤═══════════╤══════════════════════╕
+│ Total speech duration        │ 281:01:26 │ 100.00% of recording │
+├──────────────────────────────┼───────────┼──────────────────────┤
+│ Total speaking time duration │ 281:01:26 │ 100.00% of recording │
+├──────────────────────────────┼───────────┼──────────────────────┤
+│ Total silence duration       │ 00:00:00  │ 0.00% of recording   │
+╘══════════════════════════════╧═══════════╧══════════════════════╛
+
+Eval2000 Cut statistics:
+╒═══════════════════════════╤══════════╕
+│ Cuts count:               │ 4473     │
+├───────────────────────────┼──────────┤
+│ Total duration (hh:mm:ss) │ 03:37:13 │
+├───────────────────────────┼──────────┤
+│ mean                      │ 2.9      │
+├───────────────────────────┼──────────┤
+│ std                       │ 2.6      │
+├───────────────────────────┼──────────┤
+│ min                       │ 0.1      │
+├───────────────────────────┼──────────┤
+│ 25%                       │ 1.2      │
+├───────────────────────────┼──────────┤
+│ 50%                       │ 2.1      │
+├───────────────────────────┼──────────┤
+│ 75%                       │ 4.0      │
+├───────────────────────────┼──────────┤
+│ 99%                       │ 12.6     │
+├───────────────────────────┼──────────┤
+│ 99.5%                     │ 13.7     │
+├───────────────────────────┼──────────┤
+│ 99.9%                     │ 14.7     │
+├───────────────────────────┼──────────┤
+│ max                       │ 15.5     │
+├───────────────────────────┼──────────┤
+│ Recordings available:     │ 4473     │
+├───────────────────────────┼──────────┤
+│ Features available:       │ 4473     │
+├───────────────────────────┼──────────┤
+│ Supervisions available:   │ 4473     │
+╘═══════════════════════════╧══════════╛
+Speech duration statistics:
+╒══════════════════════════════╤══════════╤══════════════════════╕
+│ Total speech duration        │ 03:37:13 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total speaking time duration │ 03:37:13 │ 100.00% of recording │
+├──────────────────────────────┼──────────┼──────────────────────┤
+│ Total silence duration       │ 00:00:00 │ 0.00% of recording   │
+╘══════════════════════════════╧══════════╧══════════════════════╛
+"""
diff --git a/egs/swbd/ASR/local/extend_segments.pl b/egs/swbd/ASR/local/extend_segments.pl
new file mode 100755
index 000000000..e8b4894d5
--- /dev/null
+++ b/egs/swbd/ASR/local/extend_segments.pl
@@ -0,0 +1,99 @@
+#!/usr/bin/env perl
+use warnings; #sed replacement for -w perl parameter
+
+if (@ARGV != 1 || !($ARGV[0] =~ m/^-?\d+\.?\d*$/ &&  $ARGV[0] >= 0)) {
+  print STDERR "Usage: extend_segments.pl time-in-seconds <segments >segments.extended \n" .
+       "e.g. extend_segments.pl 0.25 <segments.1 >segments.2\n" .
+       "This command modifies a segments file, with lines like\n" .
+       " <utterance-id> <recording-id> <start-time> <end-time>\n" .
+       "by extending the beginning and end of each segment by a certain\n" .
+       "length of time.  This script makes sure the output segments do not\n" .
+       "overlap as a result of this time-extension, and that there are no\n" .
+       "negative times in the output.\n";
+  exit 1;
+}
+
+$extend = $ARGV[0];
+
+@all_lines = ();
+
+while (<STDIN>) {
+  chop;
+  @A = split(" ", $_);
+  if (@A != 4) {
+    die "invalid line in segments file: $_";
+  }
+  $line = @all_lines;  # current number of lines.
+  ($utt_id, $reco_id, $start_time, $end_time) = @A;
+
+  push @all_lines, [ $utt_id, $reco_id, $start_time, $end_time ]; # anonymous array.
+  if (! defined $lines_for_reco{$reco_id}) {
+    $lines_for_reco{$reco_id} = [ ];  # push new anonymous array.
+  }
+  push @{$lines_for_reco{$reco_id}}, $line;
+}
+
+foreach $reco_id (keys %lines_for_reco) {
+  $ref = $lines_for_reco{$reco_id};
+  @line_numbers = sort { ${$all_lines[$a]}[2] <=> ${$all_lines[$b]}[2] } @$ref;
+
+
+  {
+    # handle start of earliest segment as a special case.
+    $l0 = $line_numbers[0];
+    $tstart = ${$all_lines[$l0]}[2] - $extend;
+    if ($tstart < 0.0) { $tstart = 0.0; }
+    ${$all_lines[$l0]}[2] = $tstart;
+  }
+  {
+    # handle end of latest segment as a special case.
+    $lN = $line_numbers[$#line_numbers];
+    $tend = ${$all_lines[$lN]}[3] + $extend;
+    ${$all_lines[$lN]}[3] = $tend;
+  }
+  for ($i = 0; $i < $#line_numbers; $i++) {
+    $ln = $line_numbers[$i];
+    $ln1 = $line_numbers[$i+1];
+    $tend = ${$all_lines[$ln]}[3]; # end of earlier segment.
+    $tstart = ${$all_lines[$ln1]}[2]; # start of later segment.
+    if ($tend > $tstart) {
+      $utt1 = ${$all_lines[$ln]}[0];
+      $utt2 = ${$all_lines[$ln1]}[0];
+      print STDERR "Warning: for utterances $utt1 and $utt2, segments " .
+        "already overlap; leaving these times unchanged.\n";
+    } else {
+      $my_extend = $extend;
+      $max_extend =  0.5 * ($tstart - $tend);
+      if ($my_extend > $max_extend) { $my_extend = $max_extend; }
+      $tend += $my_extend;
+      $tstart -= $my_extend;
+      ${$all_lines[$ln]}[3] = $tend;
+      ${$all_lines[$ln1]}[2] = $tstart;
+    }
+  }
+}
+
+# leave the numbering of the lines unchanged.
+for ($l = 0; $l < @all_lines; $l++) {
+  $ref = $all_lines[$l];
+  ($utt_id, $reco_id, $start_time, $end_time) = @$ref;
+  printf("%s %s %.2f %.2f\n", $utt_id, $reco_id, $start_time, $end_time);
+}
+
+__END__
+
+# testing below.
+
+# ( echo a1 A 0 1; echo a2 A 3 4; echo b1 B 0 1; echo b2 B 2 3 ) | local/extend_segments.pl 1.0
+a1 A 0.00 2.00
+a2 A 2.00 5.00
+b1 B 0.00 1.50
+b2 B 1.50 4.00
+# ( echo a1 A 0 2; echo a2 A 1 3 ) | local/extend_segments.pl 1.0
+Warning: for utterances a1 and a2, segments already overlap; leaving these times unchanged.
+a1 A 0.00 2.00
+a2 A 1.00 4.00
+# ( echo a1 A 0 2; echo a2 A 5 6; echo a3 A 3 4 ) | local/extend_segments.pl 1.0
+a1 A 0.00 2.50
+a2 A 4.50 7.00
+a3 A 2.50 4.50
diff --git a/egs/swbd/ASR/local/filter_cuts.py b/egs/swbd/ASR/local/filter_cuts.py
new file mode 100755
index 000000000..fbcc9e24a
--- /dev/null
+++ b/egs/swbd/ASR/local/filter_cuts.py
@@ -0,0 +1,160 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This script removes short and long utterances from a cutset.
+
+Caution:
+  You may need to tune the thresholds for your own dataset.
+
+Usage example:
+
+  python3 ./local/filter_cuts.py \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --in-cuts data/fbank/librispeech_cuts_test-clean.jsonl.gz \
+    --out-cuts data/fbank-filtered/librispeech_cuts_test-clean.jsonl.gz
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+from lhotse import CutSet, load_manifest_lazy
+from lhotse.cut import Cut
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--bpe-model",
+        type=Path,
+        help="Path to the bpe.model",
+    )
+
+    parser.add_argument(
+        "--in-cuts",
+        type=Path,
+        help="Path to the input cutset",
+    )
+
+    parser.add_argument(
+        "--out-cuts",
+        type=Path,
+        help="Path to the output cutset",
+    )
+
+    return parser.parse_args()
+
+
+def filter_cuts(cut_set: CutSet, sp: spm.SentencePieceProcessor):
+    total = 0  # number of total utterances before removal
+    removed = 0  # number of removed utterances
+
+    def remove_short_and_long_utterances(c: Cut):
+        """Return False to exclude the input cut"""
+        nonlocal removed, total
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ./display_manifest_statistics.py
+        #
+        # You should use ./display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        total += 1
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            removed += 1
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./pruned_transducer_stateless2/conformer.py, the
+        # conv module uses the following expression
+        # for subsampling
+        if c.num_frames is None:
+            num_frames = c.duration * 100  # approximate
+        else:
+            num_frames = c.num_frames
+
+        T = ((num_frames - 1) // 2 - 1) // 2
+        # Note: for ./lstm_transducer_stateless/lstm.py, the formula is
+        #  T = ((num_frames - 3) // 2 - 1) // 2
+
+        # Note: for ./pruned_transducer_stateless7/zipformer.py, the formula is
+        # T = ((num_frames - 7) // 2 + 1) // 2
+
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            removed += 1
+            return False
+
+        return True
+
+    # We use to_eager() here so that we can print out the value of total
+    # and removed below.
+    ans = cut_set.filter(remove_short_and_long_utterances).to_eager()
+    ratio = removed / total * 100
+    logging.info(
+        f"Removed {removed} cuts from {total} cuts. {ratio:.3f}% data is removed."
+    )
+    return ans
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    if args.out_cuts.is_file():
+        logging.info(f"{args.out_cuts} already exists - skipping")
+        return
+
+    assert args.in_cuts.is_file(), f"{args.in_cuts} does not exist"
+    assert args.bpe_model.is_file(), f"{args.bpe_model} does not exist"
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(args.bpe_model))
+
+    cut_set = load_manifest_lazy(args.in_cuts)
+    assert isinstance(cut_set, CutSet)
+
+    cut_set = filter_cuts(cut_set, sp)
+    logging.info(f"Saving to {args.out_cuts}")
+    args.out_cuts.parent.mkdir(parents=True, exist_ok=True)
+    cut_set.to_file(args.out_cuts)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/swbd/ASR/local/filter_empty_text.py b/egs/swbd/ASR/local/filter_empty_text.py
new file mode 100755
index 000000000..6b3316800
--- /dev/null
+++ b/egs/swbd/ASR/local/filter_empty_text.py
@@ -0,0 +1,72 @@
+#!/usr/bin/env python3
+# Copyright    2023  The Chinese University of Hong Kong (author: Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+from pathlib import Path
+import logging
+from typing import List
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--kaldi-data-dir",
+        type=Path,
+        required=True,
+        help="Path to the kaldi data dir",
+    )
+
+    return parser.parse_args()
+
+
+def load_segments(path: Path):
+    segments = {}
+    with open(path, "r") as f:
+        lines = f.readlines()
+        for line in lines:
+            line = line.strip()
+            utt_id, rec_id, start, end = line.split()
+            segments[utt_id] = line
+    return segments
+
+
+def filter_text(path: Path):
+    with open(path, "r") as f:
+        lines = f.readlines()
+        return list(filter(lambda x: len(x.strip().split()) > 1, lines))
+
+
+def write_segments(path: Path, texts: List[str]):
+    with open(path, "w") as f:
+        f.writelines(texts)
+
+
+def main():
+    args = get_args()
+    orig_text_dict = filter_text(args.kaldi_data_dir / "text")
+    write_segments(args.kaldi_data_dir / "text", orig_text_dict)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
+
+    logging.info("Empty lines filtered")
diff --git a/egs/swbd/ASR/local/format_acronyms_dict.py b/egs/swbd/ASR/local/format_acronyms_dict.py
new file mode 100755
index 000000000..fa598dd03
--- /dev/null
+++ b/egs/swbd/ASR/local/format_acronyms_dict.py
@@ -0,0 +1,118 @@
+#!/usr/bin/env python3
+
+# Copyright 2015  Minhua Wu
+# Apache 2.0
+
+# convert acronyms in swbd dict to fisher convention
+# IBM to i._b._m.
+# BBC to b._b._c.
+# BBCs to b._b._c.s
+# BBC's to b._b._c.'s
+
+import argparse
+import re
+
+__author__ = "Minhua Wu"
+
+parser = argparse.ArgumentParser(description="format acronyms to a._b._c.")
+parser.add_argument("-i", "--input", help="Input lexicon", required=True)
+parser.add_argument("-o", "--output", help="Output lexicon", required=True)
+parser.add_argument(
+    "-L", "--Letter", help="Input single letter pronunciation", required=True
+)
+parser.add_argument("-M", "--Map", help="Output acronyms mapping", required=True)
+args = parser.parse_args()
+
+
+fin_lex = open(args.input, "r")
+fin_Letter = open(args.Letter, "r")
+fout_lex = open(args.output, "w")
+fout_map = open(args.Map, "w")
+
+# Initialise single letter dictionary
+dict_letter = {}
+for single_letter_lex in fin_Letter:
+    items = single_letter_lex.split()
+    dict_letter[items[0]] = single_letter_lex[len(items[0]) + 1 :].strip()
+fin_Letter.close()
+# print dict_letter
+
+for lex in fin_lex:
+    items = lex.split()
+    word = items[0]
+    lexicon = lex[len(items[0]) + 1 :].strip()
+    # find acronyms from words with only letters and '
+    pre_match = re.match(r"^[A-Za-z]+$|^[A-Za-z]+\'s$|^[A-Za-z]+s$", word)
+    if pre_match:
+        # find if words in the form of xxx's is acronym
+        if word[-2:] == "'s" and (lexicon[-1] == "s" or lexicon[-1] == "z"):
+            actual_word = word[:-2]
+            actual_lexicon = lexicon[:-2]
+            acronym_lexicon = ""
+            for w in actual_word:
+                acronym_lexicon = acronym_lexicon + dict_letter[w.upper()] + " "
+            if acronym_lexicon.strip() == actual_lexicon:
+                acronym_mapped = ""
+                acronym_mapped_back = ""
+                for w in actual_word[:-1]:
+                    acronym_mapped = acronym_mapped + w.lower() + "._"
+                    acronym_mapped_back = acronym_mapped_back + w.lower() + " "
+                acronym_mapped = acronym_mapped + actual_word[-1].lower() + ".'s"
+                acronym_mapped_back = (
+                    acronym_mapped_back + actual_word[-1].lower() + "'s"
+                )
+                fout_map.write(
+                    word + "\t" + acronym_mapped + "\t" + acronym_mapped_back + "\n"
+                )
+                fout_lex.write(acronym_mapped + " " + lexicon + "\n")
+            else:
+                fout_lex.write(lex)
+
+        # find if words in the form of xxxs is acronym
+        elif word[-1] == "s" and (lexicon[-1] == "s" or lexicon[-1] == "z"):
+            actual_word = word[:-1]
+            actual_lexicon = lexicon[:-2]
+            acronym_lexicon = ""
+            for w in actual_word:
+                acronym_lexicon = acronym_lexicon + dict_letter[w.upper()] + " "
+            if acronym_lexicon.strip() == actual_lexicon:
+                acronym_mapped = ""
+                acronym_mapped_back = ""
+                for w in actual_word[:-1]:
+                    acronym_mapped = acronym_mapped + w.lower() + "._"
+                    acronym_mapped_back = acronym_mapped_back + w.lower() + " "
+                acronym_mapped = acronym_mapped + actual_word[-1].lower() + ".s"
+                acronym_mapped_back = (
+                    acronym_mapped_back + actual_word[-1].lower() + "'s"
+                )
+                fout_map.write(
+                    word + "\t" + acronym_mapped + "\t" + acronym_mapped_back + "\n"
+                )
+                fout_lex.write(acronym_mapped + " " + lexicon + "\n")
+            else:
+                fout_lex.write(lex)
+
+        # find if words in the form of xxx (not ended with 's or s) is acronym
+        elif word.find("'") == -1 and word[-1] != "s":
+            acronym_lexicon = ""
+            for w in word:
+                acronym_lexicon = acronym_lexicon + dict_letter[w.upper()] + " "
+            if acronym_lexicon.strip() == lexicon:
+                acronym_mapped = ""
+                acronym_mapped_back = ""
+                for w in word[:-1]:
+                    acronym_mapped = acronym_mapped + w.lower() + "._"
+                    acronym_mapped_back = acronym_mapped_back + w.lower() + " "
+                acronym_mapped = acronym_mapped + word[-1].lower() + "."
+                acronym_mapped_back = acronym_mapped_back + word[-1].lower()
+                fout_map.write(
+                    word + "\t" + acronym_mapped + "\t" + acronym_mapped_back + "\n"
+                )
+                fout_lex.write(acronym_mapped + " " + lexicon + "\n")
+            else:
+                fout_lex.write(lex)
+        else:
+            fout_lex.write(lex)
+
+    else:
+        fout_lex.write(lex)
diff --git a/egs/swbd/ASR/local/generate_unique_lexicon.py b/egs/swbd/ASR/local/generate_unique_lexicon.py
new file mode 100755
index 000000000..3459c2f5a
--- /dev/null
+++ b/egs/swbd/ASR/local/generate_unique_lexicon.py
@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file takes as input a lexicon.txt and output a new lexicon,
+in which each word has a unique pronunciation.
+
+The way to do this is to keep only the first pronunciation of a word
+in lexicon.txt.
+"""
+
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain a file lexicon.txt.
+        This file will generate a new file uniq_lexicon.txt
+        in it.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def filter_multiple_pronunications(
+    lexicon: List[Tuple[str, List[str]]]
+) -> List[Tuple[str, List[str]]]:
+    """Remove multiple pronunciations of words from a lexicon.
+
+    If a word has more than one pronunciation in the lexicon, only
+    the first one is kept, while other pronunciations are removed
+    from the lexicon.
+
+    Args:
+      lexicon:
+        The input lexicon, containing a list of (word, [p1, p2, ..., pn]),
+        where "p1, p2, ..., pn" are the pronunciations of the "word".
+    Returns:
+      Return a new lexicon where each word has a unique pronunciation.
+    """
+    seen = set()
+    ans = []
+
+    for word, tokens in lexicon:
+        if word in seen:
+            continue
+        seen.add(word)
+        ans.append((word, tokens))
+    return ans
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    lexicon_filename = lang_dir / "lexicon.txt"
+
+    in_lexicon = read_lexicon(lexicon_filename)
+
+    out_lexicon = filter_multiple_pronunications(in_lexicon)
+
+    write_lexicon(lang_dir / "uniq_lexicon.txt", out_lexicon)
+
+    logging.info(f"Number of entries in lexicon.txt: {len(in_lexicon)}")
+    logging.info(f"Number of entries in uniq_lexicon.txt: {len(out_lexicon)}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/swbd/ASR/local/map_acronyms_transcripts.py b/egs/swbd/ASR/local/map_acronyms_transcripts.py
new file mode 100755
index 000000000..ba02aaec3
--- /dev/null
+++ b/egs/swbd/ASR/local/map_acronyms_transcripts.py
@@ -0,0 +1,60 @@
+#!/usr/bin/env python3
+
+# Copyright 2015  Minhua Wu
+# Apache 2.0
+
+# convert acronyms in swbd transcript to fisher convention
+# according to first two columns in the input acronyms mapping
+
+import argparse
+import re
+
+__author__ = "Minhua Wu"
+
+parser = argparse.ArgumentParser(description="format acronyms to a._b._c.")
+parser.add_argument("-i", "--input", help="Input transcripts", required=True)
+parser.add_argument("-o", "--output", help="Output transcripts", required=True)
+parser.add_argument("-M", "--Map", help="Input acronyms mapping", required=True)
+args = parser.parse_args()
+
+fin_map = open(args.Map, "r")
+dict_acronym = {}
+dict_acronym_noi = {}  # Mapping of acronyms without I, i
+for pair in fin_map:
+    items = pair.split("\t")
+    dict_acronym[items[0]] = items[1]
+    dict_acronym_noi[items[0]] = items[1]
+fin_map.close()
+del dict_acronym_noi["I"]
+del dict_acronym_noi["i"]
+
+
+fin_trans = open(args.input, "r")
+fout_trans = open(args.output, "w")
+for line in fin_trans:
+    items = line.split()
+    L = len(items)
+    # First pass mapping to map I as part of acronym
+    for i in range(L):
+        if items[i] == "I":
+            x = 0
+            while i - 1 - x >= 0 and re.match(r"^[A-Z]$", items[i - 1 - x]):
+                x += 1
+
+            y = 0
+            while i + 1 + y < L and re.match(r"^[A-Z]$", items[i + 1 + y]):
+                y += 1
+
+            if x + y > 0:
+                for bias in range(-x, y + 1):
+                    items[i + bias] = dict_acronym[items[i + bias]]
+
+    # Second pass mapping (not mapping 'i' and 'I')
+    for i in range(len(items)):
+        if items[i] in dict_acronym_noi.keys():
+            items[i] = dict_acronym_noi[items[i]]
+    sentence = " ".join(items[1:])
+    fout_trans.write(items[0] + " " + sentence.lower() + "\n")
+
+fin_trans.close()
+fout_trans.close()
diff --git a/egs/swbd/ASR/local/normalize_and_filter_supervisions.py b/egs/swbd/ASR/local/normalize_and_filter_supervisions.py
new file mode 100755
index 000000000..20ab90caf
--- /dev/null
+++ b/egs/swbd/ASR/local/normalize_and_filter_supervisions.py
@@ -0,0 +1,283 @@
+#!/usr/bin/env python3
+# Copyright    2023      (authors: Nagendra Goel https://github.com/ngoel17)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import re
+from typing import Tuple
+
+from lhotse import SupervisionSegment, SupervisionSet
+from lhotse.serialization import load_manifest_lazy_or_eager
+from tqdm import tqdm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("input_sups")
+    parser.add_argument("output_sups")
+    return parser.parse_args()
+
+
+# replacement function to convert lowercase letter to uppercase
+def to_upper(match_obj):
+    if match_obj.group() is not None:
+        return match_obj.group().upper()
+
+
+def insert_groups_and_capitalize_3(match):
+    return f"{match.group(1)} {match.group(2)} {match.group(3)}".upper()
+
+
+def insert_groups_and_capitalize_2(match):
+    return f"{match.group(1)} {match.group(2)}".upper()
+
+
+def insert_groups_and_capitalize_1(match):
+    return f"{match.group(1)}".upper()
+
+
+def insert_groups_and_capitalize_1s(match):
+    return f"{match.group(1)}".upper() + "'s"
+
+
+class FisherSwbdNormalizer:
+    """Note: the functions "normalize" and "keep" implement the logic
+    similar to Kaldi's data prep scripts for Fisher and SWBD: One
+    notable difference is that we don't change [cough], [lipsmack],
+    etc. to [noise].  We also don't implement all the edge cases of
+    normalization from Kaldi (hopefully won't make too much
+    difference).
+    """
+
+    def __init__(self) -> None:
+        self.remove_regexp_before = re.compile(
+            r"|".join(
+                [
+                    # special symbols
+                    r"\[\[skip.*\]\]",
+                    r"\[skip.*\]",
+                    r"\[pause.*\]",
+                    r"\[silence\]",
+                    r"<b_aside>",
+                    r"<e_aside>",
+                    r"_1",
+                ]
+            )
+        )
+
+        # tuples of (pattern, replacement)
+        # note: Kaldi replaces sighs, coughs, etc with [noise].
+        #       We don't do that here.
+        #       We also lowercase the text as the first operation.
+        self.replace_regexps: Tuple[re.Pattern, str] = [
+            # SWBD:
+            # [LAUGHTER-STORY] -> STORY
+            (re.compile(r"\[laughter-(.*?)\]"), r"\1"),
+            # [WEA[SONABLE]-/REASONABLE]
+            (re.compile(r"\[\S+/(\S+)\]"), r"\1"),
+            # -[ADV]AN[TAGE]- -> AN
+            (re.compile(r"-?\[.*?\](\w+)\[.*?\]-?"), r"\1-"),
+            # ABSOLUTE[LY]- -> ABSOLUTE-
+            (re.compile(r"(\w+)\[.*?\]-?"), r"\1-"),
+            # [AN]Y- -> Y-
+            # -[AN]Y- -> Y-
+            (re.compile(r"-?\[.*?\](\w+)-?"), r"\1-"),
+            # special tokens
+            (re.compile(r"\[laugh.*?\]"), r"[laughter]"),
+            (re.compile(r"\[sigh.*?\]"), r"[sigh]"),
+            (re.compile(r"\[cough.*?\]"), r"[cough]"),
+            (re.compile(r"\[mn.*?\]"), r"[vocalized-noise]"),
+            (re.compile(r"\[breath.*?\]"), r"[breath]"),
+            (re.compile(r"\[lipsmack.*?\]"), r"[lipsmack]"),
+            (re.compile(r"\[sneeze.*?\]"), r"[sneeze]"),
+            # abbreviations
+            (
+                re.compile(
+                    r"(\w)\.(\w)\.(\w)",
+                ),
+                insert_groups_and_capitalize_3,
+            ),
+            (
+                re.compile(
+                    r"(\w)\.(\w)",
+                ),
+                insert_groups_and_capitalize_2,
+            ),
+            (
+                re.compile(
+                    r"([a-h,j-z])\.",
+                ),
+                insert_groups_and_capitalize_1,
+            ),
+            (
+                re.compile(
+                    r"\._",
+                ),
+                r" ",
+            ),
+            (
+                re.compile(
+                    r"_(\w)",
+                ),
+                insert_groups_and_capitalize_1,
+            ),
+            (
+                re.compile(
+                    r"(\w)\.s",
+                ),
+                insert_groups_and_capitalize_1s,
+            ),
+            (
+                re.compile(
+                    r"([A-Z])\'s",
+                ),
+                insert_groups_and_capitalize_1s,
+            ),
+            (
+                re.compile(
+                    r"(\s\w\b|^\w\b)",
+                ),
+                insert_groups_and_capitalize_1,
+            ),
+            # words between apostrophes
+            (re.compile(r"'(\S*?)'"), r"\1"),
+            # dangling dashes (2 passes)
+            (re.compile(r"\s-\s"), r" "),
+            (re.compile(r"\s-\s"), r" "),
+            # special symbol with trailing dash
+            (re.compile(r"(\[.*?\])-"), r"\1"),
+            # Just remove all dashes
+            (re.compile(r"-"), r" "),
+        ]
+
+        # unwanted symbols in the transcripts
+        self.remove_regexp_after = re.compile(
+            r"|".join(
+                [
+                    # remaining punctuation
+                    r"\.",
+                    r",",
+                    r"\?",
+                    r"{",
+                    r"}",
+                    r"~",
+                    r"_\d",
+                ]
+            )
+        )
+
+        self.post_fixes = [
+            # Fix an issue related to [VOCALIZED NOISE] after dash removal
+            (re.compile(r"\[vocalized noise\]"), "[vocalized-noise]"),
+        ]
+
+        self.whitespace_regexp = re.compile(r"\s+")
+
+    def normalize(self, text: str) -> str:
+        text = text.lower()
+
+        # first remove
+        text = self.remove_regexp_before.sub("", text)
+
+        # then replace
+        for pattern, sub in self.replace_regexps:
+            text = pattern.sub(sub, text)
+
+        # then remove
+        text = self.remove_regexp_after.sub("", text)
+
+        # post fixes
+        for pattern, sub in self.post_fixes:
+            text = pattern.sub(sub, text)
+
+        # then clean up whitespace
+        text = self.whitespace_regexp.sub(" ", text).strip()
+
+        return text.upper()
+
+
+def keep(sup: SupervisionSegment) -> bool:
+    if "((" in sup.text:
+        return False
+
+    if "<german" in sup.text:
+        return False
+
+    return True
+
+
+def main():
+    args = get_args()
+    sups = load_manifest_lazy_or_eager(args.input_sups)
+    assert isinstance(sups, SupervisionSet)
+
+    normalizer = FisherSwbdNormalizer()
+
+    tot, skip = 0, 0
+    with SupervisionSet.open_writer(args.output_sups) as writer:
+        for sup in tqdm(sups, desc="Normalizing supervisions"):
+            tot += 1
+
+            if not keep(sup):
+                skip += 1
+                continue
+
+            sup.text = normalizer.normalize(sup.text).upper()
+            if not sup.text:
+                skip += 1
+                continue
+
+            writer.write(sup)
+    print(f"tot: {tot}, skip: {skip}")
+
+
+def test():
+    normalizer = FisherSwbdNormalizer()
+    for text in [
+        "[laughterr] [SILENCE]",
+        "[laugh] oh this is great [silence] <B_ASIDE> yes",
+        "[laugh] oh this is [laught] this is great [silence] <B_ASIDE> yes",
+        "i don't kn- - know A.B.C's",
+        "so x. corp is good?",
+        "'absolutely yes",
+        "absolutely' yes",
+        "'absolutely' yes",
+        "'absolutely' yes 'aight",
+        "ABSOLUTE[LY]",
+        "ABSOLUTE[LY]-",
+        "[AN]Y",
+        "[AN]Y-",
+        "[ADV]AN[TAGE]",
+        "[ADV]AN[TAGE]-",
+        "-[ADV]AN[TAGE]",
+        "-[ADV]AN[TAGE]-",
+        "[WEA[SONABLE]-/REASONABLE]",
+        "[VOCALIZED-NOISE]-",
+        "~BULL",
+        "Frank E Peretti P E R E T T I",
+        "yeah yeah like Double O Seven he's supposed to do it",
+        "P A P E R paper",
+        "[noise] okay_1 um let me see [laughter] i've been sitting here awhile",
+    ]:
+        print(text)
+        print(normalizer.normalize(text))
+        print()
+
+
+if __name__ == "__main__":
+    test()
+    # exit()
+    main()
diff --git a/egs/swbd/ASR/local/normalize_eval2000.py b/egs/swbd/ASR/local/normalize_eval2000.py
new file mode 100755
index 000000000..7316193d0
--- /dev/null
+++ b/egs/swbd/ASR/local/normalize_eval2000.py
@@ -0,0 +1,234 @@
+#!/usr/bin/env python3
+# Copyright    2023      (authors: Nagendra Goel https://github.com/ngoel17)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import re
+from typing import Tuple
+
+from lhotse import SupervisionSegment, SupervisionSet
+from lhotse.serialization import load_manifest_lazy_or_eager
+from tqdm import tqdm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("input_sups")
+    parser.add_argument("output_sups")
+    return parser.parse_args()
+
+
+def remove_punctutation_and_other_symbol(text: str) -> str:
+    text = text.replace("--", " ")
+    text = text.replace("//", " ")
+    text = text.replace(".", " ")
+    text = text.replace("?", " ")
+    text = text.replace("~", " ")
+    text = text.replace(",", " ")
+    text = text.replace(";", " ")
+    text = text.replace("(", " ")
+    text = text.replace(")", " ")
+    text = text.replace("&", " ")
+    text = text.replace("%", " ")
+    text = text.replace("*", " ")
+    text = text.replace("{", " ")
+    text = text.replace("}", " ")
+    return text
+
+
+def eval2000_clean_eform(text: str, eform_count) -> str:
+    string_to_remove = []
+    piece = text.split('">')
+    for i in range(0, len(piece)):
+        s = piece[i] + '">'
+        res = re.search(r"<contraction e_form(.*?)\">", s)
+        if res is not None:
+            res_rm = res.group(1)
+            string_to_remove.append(res_rm)
+    for p in string_to_remove:
+        eform_string = p
+        text = text.replace(eform_string, " ")
+    eform_1 = "<contraction e_form"
+    text = text.replace(eform_1, " ")
+    eform_2 = '">'
+    text = text.replace(eform_2, " ")
+    # print("TEXT final: ", text)
+    return text
+
+
+def replace_silphone(text: str) -> str:
+    text = text.replace("[/BABY CRYING]", " ")
+    text = text.replace("[/CHILD]", " ")
+    text = text.replace("[[DISTORTED]]", " ")
+    text = text.replace("[/DISTORTION]", " ")
+    text = text.replace("[[DRAWN OUT]]", " ")
+    text = text.replace("[[DRAWN-OUT]]", " ")
+    text = text.replace("[[FAINT]]", " ")
+    text = text.replace("[SMACK]", " ")
+    text = text.replace("[[MUMBLES]]", " ")
+    text = text.replace("[[HIGH PITCHED SQUEAKY VOICE]]", " ")
+    text = text.replace("[[IN THE LAUGH]]", "[LAUGHTER]")
+    text = text.replace("[[LAST WORD SPOKEN WITH A LAUGH]]", "[LAUGHTER]")
+    text = text.replace("[[PART OF FIRST SYLLABLE OF PREVIOUS WORD CUT OFF]]", " ")
+    text = text.replace("[[PREVIOUS WORD SPOKEN WITH A LAUGH]]", " ")
+    text = text.replace("[[PREVIOUS TWO WORDS SPOKEN WHILE LAUGHING]]", " ")
+    text = text.replace("[[PROLONGED]]", " ")
+    text = text.replace("[/RUNNING WATER]", " ")
+    text = text.replace("[[SAYS LAUGHING]]", "[LAUGHTER]")
+    text = text.replace("[[SINGING]]", " ")
+    text = text.replace("[[SPOKEN WHILE LAUGHING]]", "[LAUGHTER]")
+    text = text.replace("[/STATIC]", " ")
+    text = text.replace("['THIRTIETH' DRAWN OUT]", " ")
+    text = text.replace("[/VOICES]", " ")
+    text = text.replace("[[WHISPERED]]", " ")
+    text = text.replace("[DISTORTION]", " ")
+    text = text.replace("[DISTORTION, HIGH VOLUME ON WAVES]", " ")
+    text = text.replace("[BACKGROUND LAUGHTER]", "[LAUGHTER]")
+    text = text.replace("[CHILD'S VOICE]", " ")
+    text = text.replace("[CHILD SCREAMS]", " ")
+    text = text.replace("[CHILD VOICE]", " ")
+    text = text.replace("[CHILD YELLING]", " ")
+    text = text.replace("[CHILD SCREAMING]", " ")
+    text = text.replace("[CHILD'S VOICE IN BACKGROUND]", " ")
+    text = text.replace("[CHANNEL NOISE]", " ")
+    text = text.replace("[CHANNEL ECHO]", " ")
+    text = text.replace("[ECHO FROM OTHER CHANNEL]", " ")
+    text = text.replace("[ECHO OF OTHER CHANNEL]", " ")
+    text = text.replace("[CLICK]", " ")
+    text = text.replace("[DISTORTED]", " ")
+    text = text.replace("[BABY CRYING]", " ")
+    text = text.replace("[METALLIC KNOCKING SOUND]", " ")
+    text = text.replace("[METALLIC SOUND]", " ")
+
+    text = text.replace("[PHONE JIGGLING]", " ")
+    text = text.replace("[BACKGROUND SOUND]", " ")
+    text = text.replace("[BACKGROUND VOICE]", " ")
+    text = text.replace("[BACKGROUND VOICES]", " ")
+    text = text.replace("[BACKGROUND NOISE]", " ")
+    text = text.replace("[CAR HORNS IN BACKGROUND]", " ")
+    text = text.replace("[CAR HORNS]", " ")
+    text = text.replace("[CARNATING]", " ")
+    text = text.replace("[CRYING CHILD]", " ")
+    text = text.replace("[CHOPPING SOUND]", " ")
+    text = text.replace("[BANGING]", " ")
+    text = text.replace("[CLICKING NOISE]", " ")
+    text = text.replace("[CLATTERING]", " ")
+    text = text.replace("[ECHO]", " ")
+    text = text.replace("[KNOCK]", " ")
+    text = text.replace("[NOISE-GOOD]", "[NOISE]")
+    text = text.replace("[RIGHT]", " ")
+    text = text.replace("[SOUND]", " ")
+    text = text.replace("[SQUEAK]", " ")
+    text = text.replace("[STATIC]", " ")
+    text = text.replace("[[SAYS WITH HIGH-PITCHED SCREAMING LAUGHTER]]", " ")
+    text = text.replace("[UH]", "UH")
+    text = text.replace("[MN]", "[VOCALIZED-NOISE]")
+    text = text.replace("[VOICES]", " ")
+    text = text.replace("[WATER RUNNING]", " ")
+    text = text.replace("[SOUND OF TWISTING PHONE CORD]", " ")
+    text = text.replace("[SOUND OF SOMETHING FALLING]", " ")
+    text = text.replace("[SOUND]", " ")
+    text = text.replace("[NOISE OF MOVING PHONE]", " ")
+    text = text.replace("[SOUND OF RUNNING WATER]", " ")
+    text = text.replace("[CHANNEL]", " ")
+    text = text.replace("[SILENCE]", " ")
+    text = text.replace("-[W]HERE", "WHERE")
+    text = text.replace("Y[OU]I-", "YOU I")
+    text = text.replace("-[A]ND", "AND")
+    text = text.replace("JU[ST]", "JUST")
+    text = text.replace("{BREATH}", " ")
+    text = text.replace("{BREATHY}", " ")
+    text = text.replace("{CHANNEL NOISE}", " ")
+    text = text.replace("{CLEAR THROAT}", " ")
+
+    text = text.replace("{CLEARING THROAT}", " ")
+    text = text.replace("{CLEARS THROAT}", " ")
+    text = text.replace("{COUGH}", " ")
+    text = text.replace("{DRAWN OUT}", " ")
+    text = text.replace("{EXHALATION}", " ")
+    text = text.replace("{EXHALE}", " ")
+    text = text.replace("{GASP}", " ")
+    text = text.replace("{HIGH SQUEAL}", " ")
+    text = text.replace("{INHALE}", " ")
+    text = text.replace("{LAUGH}", "[LAUGHTER]")
+    text = text.replace("{LAUGH}", "[LAUGHTER]")
+    text = text.replace("{LAUGH}", "[LAUGHTER]")
+    text = text.replace("{LIPSMACK}", " ")
+    text = text.replace("{LIPSMACK}", " ")
+
+    text = text.replace("{NOISE OF DISGUST}", " ")
+    text = text.replace("{SIGH}", " ")
+    text = text.replace("{SNIFF}", " ")
+    text = text.replace("{SNORT}", " ")
+    text = text.replace("{SHARP EXHALATION}", " ")
+    text = text.replace("{BREATH LAUGH}", " ")
+
+    text = text.replace("[LAUGHTER]", " ")
+    text = text.replace("[NOISE]", " ")
+    text = text.replace("[VOCALIZED-NOISE]", " ")
+    text = text.replace("-", " ")
+    return text
+
+
+def remove_languagetag(text: str) -> str:
+    langtag = re.findall(r"<(.*?)>", text)
+    for t in langtag:
+        text = text.replace(t, " ")
+    text = text.replace("<", " ")
+    text = text.replace(">", " ")
+    return text
+
+
+def eval2000_normalizer(text: str) -> str:
+    # print("TEXT original: ",text)
+    eform_count = text.count("contraction e_form")
+    # print("eform corunt:", eform_count)
+    if eform_count > 0:
+        text = eval2000_clean_eform(text, eform_count)
+    text = text.upper()
+    text = remove_languagetag(text)
+    text = replace_silphone(text)
+    text = remove_punctutation_and_other_symbol(text)
+    text = text.replace("IGNORE_TIME_SEGMENT_IN_SCORING", " ")
+    text = text.replace("IGNORE_TIME_SEGMENT_SCORING", " ")
+    spaces = re.findall(r"\s+", text)
+    for sp in spaces:
+        text = text.replace(sp, " ")
+    text = text.strip()
+    # text = self.whitespace_regexp.sub(" ", text).strip()
+    # print(text)
+    return text
+
+
+def main():
+    args = get_args()
+    sups = load_manifest_lazy_or_eager(args.input_sups)
+    assert isinstance(sups, SupervisionSet)
+
+    tot, skip = 0, 0
+    with SupervisionSet.open_writer(args.output_sups) as writer:
+        for sup in tqdm(sups, desc="Normalizing supervisions"):
+            tot += 1
+            sup.text = eval2000_normalizer(sup.text)
+            if not sup.text:
+                skip += 1
+                continue
+            writer.write(sup)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/local/prepare_lang.py b/egs/swbd/ASR/local/prepare_lang.py
new file mode 120000
index 000000000..747f2ab39
--- /dev/null
+++ b/egs/swbd/ASR/local/prepare_lang.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/local/prepare_lang_bpe.py b/egs/swbd/ASR/local/prepare_lang_bpe.py
new file mode 100755
index 000000000..d82a085ec
--- /dev/null
+++ b/egs/swbd/ASR/local/prepare_lang_bpe.py
@@ -0,0 +1,274 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/bpe.model,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+from icefall.utils import str2bool
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def generate_lexicon(
+    model_file: str, words: List[str], oov: str
+) -> Tuple[Lexicon, Dict[str, int]]:
+    """Generate a lexicon from a BPE model.
+
+    Args:
+      model_file:
+        Path to a sentencepiece model.
+      words:
+        A list of strings representing words.
+      oov:
+        The out of vocabulary word in lexicon.
+    Returns:
+      Return a tuple with two elements:
+        - A dict whose keys are words and values are the corresponding
+          word pieces.
+        - A dict representing the token symbol, mapping from tokens to IDs.
+    """
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(model_file))
+
+    # Convert word to word piece IDs instead of word piece strings
+    # to avoid OOV tokens.
+    words_pieces_ids: List[List[int]] = sp.encode(words, out_type=int)
+
+    # Now convert word piece IDs back to word piece strings.
+    words_pieces: List[List[str]] = [sp.id_to_piece(ids) for ids in words_pieces_ids]
+
+    lexicon = []
+    for word, pieces in zip(words, words_pieces):
+        lexicon.append((word, pieces))
+
+    lexicon.append((oov, ["▁", sp.id_to_piece(sp.unk_id())]))
+
+    token2id: Dict[str, int] = {sp.id_to_piece(i): i for i in range(sp.vocab_size())}
+
+    return lexicon, token2id
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain the bpe.model and words.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--oov",
+        type=str,
+        default="<UNK>",
+        help="The out of vocabulary word in lexicon.",
+    )
+
+    parser.add_argument(
+        "--debug",
+        type=str2bool,
+        default=False,
+        help="""True for debugging, which will generate
+        a visualization of the lexicon FST.
+
+        Caution: If your lexicon contains hundreds of thousands
+        of lines, please set it to False!
+
+        See "test/test_bpe_lexicon.py" for usage.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    model_file = lang_dir / "bpe.model"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = [
+        "<eps>",
+        "!SIL",
+        "<SPOKEN_NOISE>",
+        args.oov,
+        "#0",
+        "<s>",
+        "</s>",
+    ]
+
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    lexicon, token_sym_table = generate_lexicon(model_file, words, args.oov)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+    if args.debug:
+        labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+        L.labels_sym = labels_sym
+        L.aux_labels_sym = aux_labels_sym
+        L.draw(f"{lang_dir / 'L.svg'}", title="L.pt")
+
+        L_disambig.labels_sym = labels_sym
+        L_disambig.aux_labels_sym = aux_labels_sym
+        L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/local/prepare_lm_training_data.py b/egs/swbd/ASR/local/prepare_lm_training_data.py
new file mode 120000
index 000000000..abc00d421
--- /dev/null
+++ b/egs/swbd/ASR/local/prepare_lm_training_data.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lm_training_data.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/local/rt03_data_prep.sh b/egs/swbd/ASR/local/rt03_data_prep.sh
new file mode 100755
index 000000000..8a5f64324
--- /dev/null
+++ b/egs/swbd/ASR/local/rt03_data_prep.sh
@@ -0,0 +1,107 @@
+#!/usr/bin/env bash
+
+# RT-03 data preparation (conversational telephone speech part only)
+# Adapted from Arnab Ghoshal's script for Hub-5 Eval 2000 by Peng Qi
+
+# To be run from one directory above this script.
+
+# Expects the standard directory layout for RT-03
+
+if [ $# -ne 1 ]; then
+    echo "Usage: $0 <rt03-dir>"
+    echo "e.g.: $0 /export/corpora/LDC/LDC2007S10"
+    echo "See comments in the script for more details"
+    exit 1
+fi
+
+sdir=$1
+[ ! -d $sdir/data/audio/eval03/english/cts ] &&
+    echo Expecting directory $sdir/data/audio/eval03/english/cts to be present && exit 1
+[ ! -d $sdir/data/references/eval03/english/cts ] &&
+    echo Expecting directory $tdir/data/references/eval03/english/cts to be present && exit 1
+
+dir=data/local/rt03
+mkdir -p $dir
+
+rtroot=$sdir
+tdir=$sdir/data/references/eval03/english/cts
+sdir=$sdir/data/audio/eval03/english/cts
+
+find -L $sdir -iname '*.sph' | sort >$dir/sph.flist
+sed -e 's?.*/??' -e 's?.sph??' $dir/sph.flist | paste - $dir/sph.flist \
+    >$dir/sph.scp
+
+sph2pipe=sph2pipe
+! command -v "${sph2pipe}" &>/dev/null &&
+    echo "Could not execute the sph2pipe program at $sph2pipe" && exit 1
+
+awk -v sph2pipe=$sph2pipe '{
+    printf("%s-A %s -f wav -p -c 1 %s |\n", $1, sph2pipe, $2);
+    printf("%s-B %s -f wav -p -c 2 %s |\n", $1, sph2pipe, $2);
+}' <$dir/sph.scp | sort >$dir/wav.scp || exit 1
+#side A - channel 1, side B - channel 2
+
+# Get segments file...
+# segments file format is: utt-id side-id start-time end-time, e.g.:
+# sw02001-A_000098-001156 sw02001-A 0.98 11.56
+#pem=$sdir/english/hub5e_00.pem
+#[ ! -f $pem ] && echo "No such file $pem" && exit 1;
+# pem file has lines like:
+# en_4156 A unknown_speaker 301.85 302.48
+
+#grep -v ';;' $pem \
+cat $tdir/*.stm | grep -v ';;' | grep -v inter_segment_gap |
+    awk '{
+    spk=$1"-"(($2==1)?"A":"B");
+    utt=sprintf("%s_%06d-%06d",spk,$4*100,$5*100);
+    print utt,spk,$4,$5;}' |
+    sort -u >$dir/segments
+
+# stm file has lines like:
+# en_4156 A en_4156_A 357.64 359.64 <O,en,F,en-F>  HE IS A POLICE OFFICER
+# TODO(arnab): We should really be lowercasing this since the Edinburgh
+# recipe uses lowercase. This is not used in the actual scoring.
+#grep -v ';;' $tdir/reference/hub5e00.english.000405.stm \
+cat $tdir/*.stm | grep -v ';;' | grep -v inter_segment_gap |
+    awk '{
+    spk=$1"-"(($2==1)?"A":"B");
+    utt=sprintf("%s_%06d-%06d",spk,$4*100,$5*100);
+    printf utt; for(n=7;n<=NF;n++) printf(" %s", $n); print ""; }' |
+    sort >$dir/text.all
+
+# We'll use the stm file for sclite scoring.  There seem to be various errors
+# in the stm file that upset hubscr.pl, and we fix them here.
+cat $tdir/*.stm |
+    sed -e 's:((:(:' -e 's:<B_ASIDE>::g' -e 's:<E_ASIDE>::g' |
+    grep -v inter_segment_gap |
+    awk '{
+    printf $1; if ($1==";;") printf(" %s",$2); else printf(($2==1)?" A":" B"); for(n=3;n<=NF;n++) printf(" %s", $n); print ""; }' \
+        >$dir/stm
+#$tdir/reference/hub5e00.english.000405.stm >  $dir/stm
+cp $rtroot/data/trans_rules/en20030506.glm $dir/glm
+
+# next line uses command substitution
+# Just checking that the segments are the same in pem vs. stm.
+! cmp <(awk '{print $1}' $dir/text.all) <(awk '{print $1}' $dir/segments) &&
+    echo "Segments from pem file and stm file do not match." && exit 1
+
+grep -v IGNORE_TIME_SEGMENT_ $dir/text.all >$dir/text
+
+# create an utt2spk file that assumes each conversation side is
+# a separate speaker.
+awk '{print $1,$2;}' $dir/segments >$dir/utt2spk
+utils/utt2spk_to_spk2utt.pl $dir/utt2spk >$dir/spk2utt
+
+# cp $dir/segments $dir/segments.tmp
+# awk '{x=$3-0.05; if (x<0.0) x=0.0; y=$4+0.05; print $1, $2, x, y; }' \
+#   $dir/segments.tmp > $dir/segments
+
+awk '{print $1}' $dir/wav.scp |
+    perl -ane '$_ =~ m:^(\S+)-([AB])$: || die "bad label $_";
+    print "$1-$2 $1 $2\n"; ' \
+        >$dir/reco2file_and_channel || exit 1
+
+./utils/fix_data_dir.sh $dir
+
+echo Data preparation and formatting completed for RT-03
+echo "(but not MFCC extraction)"
diff --git a/egs/swbd/ASR/local/sort_lm_training_data.py b/egs/swbd/ASR/local/sort_lm_training_data.py
new file mode 100755
index 000000000..bed3856e4
--- /dev/null
+++ b/egs/swbd/ASR/local/sort_lm_training_data.py
@@ -0,0 +1,141 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file takes as input the filename of LM training data
+generated by ./local/prepare_lm_training_data.py and sorts
+it by sentence length.
+
+Sentence length equals to the number of BPE tokens in a sentence.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import numpy as np
+import torch
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--in-lm-data",
+        type=str,
+        help="Input LM training data, e.g., data/bpe_500/lm_data.pt",
+    )
+
+    parser.add_argument(
+        "--out-lm-data",
+        type=str,
+        help="Input LM training data, e.g., data/bpe_500/sorted_lm_data.pt",
+    )
+
+    parser.add_argument(
+        "--out-statistics",
+        type=str,
+        help="Statistics about LM training data., data/bpe_500/statistics.txt",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    in_lm_data = Path(args.in_lm_data)
+    out_lm_data = Path(args.out_lm_data)
+    assert in_lm_data.is_file(), f"{in_lm_data}"
+    if out_lm_data.is_file():
+        logging.warning(f"{out_lm_data} exists - skipping")
+        return
+    data = torch.load(in_lm_data)
+    words2bpe = data["words"]
+    sentences = data["sentences"]
+    sentence_lengths = data["sentence_lengths"]
+
+    num_sentences = sentences.dim0
+    assert num_sentences == sentence_lengths.numel(), (
+        num_sentences,
+        sentence_lengths.numel(),
+    )
+
+    indices = torch.argsort(sentence_lengths, descending=True)
+
+    sorted_sentences = sentences[indices.to(torch.int32)]
+    sorted_sentence_lengths = sentence_lengths[indices]
+
+    # Check that sentences are ordered by length
+    assert num_sentences == sorted_sentences.dim0, (
+        num_sentences,
+        sorted_sentences.dim0,
+    )
+
+    cur = None
+    for i in range(num_sentences):
+        word_ids = sorted_sentences[i]
+        token_ids = words2bpe[word_ids]
+        if isinstance(token_ids, k2.RaggedTensor):
+            token_ids = token_ids.values
+        if cur is not None:
+            assert cur >= token_ids.numel(), (cur, token_ids.numel())
+
+        cur = token_ids.numel()
+        assert cur == sorted_sentence_lengths[i]
+
+    data["sentences"] = sorted_sentences
+    data["sentence_lengths"] = sorted_sentence_lengths
+    torch.save(data, args.out_lm_data)
+    logging.info(f"Saved to {args.out_lm_data}")
+
+    statistics = Path(args.out_statistics)
+
+    # Write statistics
+    num_words = sorted_sentences.numel()
+    num_tokens = sentence_lengths.sum().item()
+    max_sentence_length = sentence_lengths[indices[0]]
+    min_sentence_length = sentence_lengths[indices[-1]]
+
+    step = 10
+    hist, bins = np.histogram(
+        sentence_lengths.numpy(),
+        bins=np.arange(1, max_sentence_length + step, step),
+    )
+
+    histogram = np.stack((bins[:-1], hist)).transpose()
+
+    with open(statistics, "w") as f:
+        f.write(f"num_sentences: {num_sentences}\n")
+        f.write(f"num_words: {num_words}\n")
+        f.write(f"num_tokens: {num_tokens}\n")
+        f.write(f"max_sentence_length: {max_sentence_length}\n")
+        f.write(f"min_sentence_length: {min_sentence_length}\n")
+        f.write("histogram:\n")
+        f.write("  bin  count  percent\n")
+        for row in histogram:
+            f.write(
+                f"{int(row[0]):>5} {int(row[1]):>5}   "
+                f"{100.*row[1]/num_sentences:.3f}%\n"
+            )
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/swbd/ASR/local/swbd1_data_prep.sh b/egs/swbd/ASR/local/swbd1_data_prep.sh
new file mode 100755
index 000000000..159359491
--- /dev/null
+++ b/egs/swbd/ASR/local/swbd1_data_prep.sh
@@ -0,0 +1,128 @@
+#!/usr/bin/env bash
+
+# Switchboard-1 training data preparation customized for Edinburgh
+# Author:  Arnab Ghoshal (Jan 2013)
+
+# To be run from one directory above this script.
+
+## The input is some directory containing the switchboard-1 release 2
+## corpus (LDC97S62).  Note: we don't make many assumptions about how
+## you unpacked this.  We are just doing a "find" command to locate
+## the .sph files.
+
+## The second input is optional, which should point to a directory containing
+## Switchboard transcriptions/documentations (specifically, the conv.tab file).
+## If specified, the script will try to use the actual speaker PINs provided
+## with the corpus instead of the conversation side ID (Kaldi default). We
+## will be using "find" to locate this file so we don't make any assumptions
+## on the directory structure. (Peng Qi, Aug 2014)
+
+#check existing directories
+if [ $# != 1 -a $# != 2 ]; then
+    echo "Usage: swbd1_data_prep.sh /path/to/SWBD [/path/to/SWBD_DOC]"
+    exit 1
+fi
+
+SWBD_DIR=$1
+
+dir=data/local/train
+mkdir -p $dir
+
+# Audio data directory check
+if [ ! -d $SWBD_DIR ]; then
+    echo "Error: run.sh requires a directory argument"
+    exit 1
+fi
+
+sph2pipe=sph2pipe
+! command -v "${sph2pipe}" &>/dev/null &&
+    echo "Could not execute the sph2pipe program at $sph2pipe" && exit 1
+
+# Option A: SWBD dictionary file check
+[ ! -f ./swb_ms98_transcriptions/sw-ms98-dict.text ] &&
+    echo "SWBD dictionary file does not exist" && exit 1
+
+# find sph audio files
+find -L $SWBD_DIR -iname '*.sph' | sort >$dir/sph.flist
+
+n=$(cat $dir/sph.flist | wc -l)
+[ $n -ne 2435 ] && [ $n -ne 2438 ] &&
+    echo Warning: expected 2435 or 2438 data data files, found $n
+
+# (1a) Transcriptions preparation
+# make basic transcription file (add segments info)
+# **NOTE: In the default Kaldi recipe, everything is made uppercase, while we
+# make everything lowercase here. This is because we will be using SRILM which
+# can optionally make everything lowercase (but not uppercase) when mapping
+# LM vocabs.
+awk '{
+name=substr($1,1,6); gsub("^sw","sw0",name); side=substr($1,7,1);
+stime=$2; etime=$3;
+printf("%s-%s_%06.0f-%06.0f",
+name, side, int(100*stime+0.5), int(100*etime+0.5));
+for(i=4;i<=NF;i++) printf(" %s", $i); printf "\n"
+}' ./swb_ms98_transcriptions/*/*/*-trans.text >$dir/transcripts1.txt
+
+# test if trans. file is sorted
+export LC_ALL=C
+sort -c $dir/transcripts1.txt || exit 1 # check it's sorted.
+
+# Remove SILENCE, <B_ASIDE> and <E_ASIDE>.
+
+# Note: we have [NOISE], [VOCALIZED-NOISE], [LAUGHTER], [SILENCE].
+# removing [SILENCE], and the <B_ASIDE> and <E_ASIDE> markers that mark
+# speech to somone; we will give phones to the other three (NSN, SPN, LAU).
+# There will also be a silence phone, SIL.
+# **NOTE: modified the pattern matches to make them case insensitive
+cat $dir/transcripts1.txt |
+    perl -ane 's:\s\[SILENCE\](\s|$):$1:gi;
+    s/<B_ASIDE>//gi;
+    s/<E_ASIDE>//gi;
+    print;' |
+    awk '{if(NF > 1) { print; } } ' >$dir/transcripts2.txt
+
+# **NOTE: swbd1_map_words.pl has been modified to make the pattern matches
+# case insensitive
+local/swbd1_map_words.pl -f 2- $dir/transcripts2.txt >$dir/text
+
+# format acronyms in text
+python3 local/map_acronyms_transcripts.py -i $dir/text -o $dir/text_map \
+    -M data/local/dict_nosp/acronyms.map
+mv $dir/text_map $dir/text
+
+# (1c) Make segment files from transcript
+#segments file format is: utt-id side-id start-time end-time, e.g.:
+#sw02001-A_000098-001156 sw02001-A 0.98 11.56
+awk '{
+segment=$1;
+split(segment,S,"[_-]");
+side=S[2]; audioname=S[1]; startf=S[3]; endf=S[4];
+print segment " " audioname "-" side " " startf/100 " " endf/100
+}' <$dir/text >$dir/segments
+
+sed -e 's?.*/??' -e 's?.sph??' $dir/sph.flist | paste - $dir/sph.flist \
+    >$dir/sph.scp
+
+awk -v sph2pipe=$sph2pipe '{
+printf("%s-A %s -f wav -p -c 1 %s |\n", $1, sph2pipe, $2);
+printf("%s-B %s -f wav -p -c 2 %s |\n", $1, sph2pipe, $2);
+}' <$dir/sph.scp | sort >$dir/wav.scp || exit 1
+#side A - channel 1, side B - channel 2
+
+# this file reco2file_and_channel maps recording-id (e.g. sw02001-A)
+# to the file name sw02001 and the A, e.g.
+# sw02001-A  sw02001 A
+# In this case it's trivial, but in other corpora the information might
+# be less obvious.  Later it will be needed for ctm scoring.
+awk '{print $1}' $dir/wav.scp |
+    perl -ane '$_ =~ m:^(\S+)-([AB])$: || die "bad label $_";
+    print "$1-$2 $1 $2\n"; ' \
+        >$dir/reco2file_and_channel || exit 1
+
+awk '{spk=substr($1,1,9); print $1 " " spk}' $dir/segments >$dir/utt2spk ||
+    exit 1
+sort -k 2 $dir/utt2spk | utils/utt2spk_to_spk2utt.pl >$dir/spk2utt || exit 1
+
+echo Switchboard-1 data preparation succeeded.
+
+utils/fix_data_dir.sh data/local/train
diff --git a/egs/swbd/ASR/local/swbd1_map_words.pl b/egs/swbd/ASR/local/swbd1_map_words.pl
new file mode 100755
index 000000000..4fb8d4ffe
--- /dev/null
+++ b/egs/swbd/ASR/local/swbd1_map_words.pl
@@ -0,0 +1,52 @@
+#!/usr/bin/env perl
+
+# Modified from swbd_map_words.pl in Kaldi s5 recipe to make pattern
+# matches case-insensitive --Arnab (Jan 2013)
+
+if ($ARGV[0] eq "-f") {
+  shift @ARGV;
+  $field_spec = shift @ARGV;
+  if ($field_spec =~ m/^\d+$/) {
+    $field_begin = $field_spec - 1; $field_end = $field_spec - 1;
+  }
+  if ($field_spec =~ m/^(\d*)[-:](\d*)/) { # accept e.g. 1:10 as a courtesy (properly, 1-10)
+    if ($1 ne "") {
+      $field_begin = $1 - 1;    # Change to zero-based indexing.
+    }
+    if ($2 ne "") {
+      $field_end = $2 - 1;      # Change to zero-based indexing.
+    }
+  }
+  if (!defined $field_begin && !defined $field_end) {
+    die "Bad argument to -f option: $field_spec";
+  }
+}
+
+
+while (<>) {
+  @A = split(" ", $_);
+  for ($n = 0; $n < @A; $n++) {
+    $a = $A[$n];
+    if ( (!defined $field_begin || $n >= $field_begin)
+         && (!defined $field_end || $n <= $field_end)) {
+      # e.g. [LAUGHTER-STORY] -> STORY;
+      $a =~ s:(|\-)^\[LAUGHTER-(.+)\](|\-)$:$1$2$3:i;
+      # $1 and $3 relate to preserving trailing "-"
+      $a =~ s:^\[(.+)/.+\](|\-)$:$1$2:; # e.g. [IT'N/ISN'T] -> IT'N ... note,
+      # 1st part may include partial-word stuff, which we process further below,
+      # e.g. [LEM[GUINI]-/LINGUINI]
+      # the (|\_) at the end is to accept and preserve trailing -'s.
+      $a =~ s:^(|\-)\[[^][]+\](.+)$:-$2:; # e.g. -[AN]Y , note \047 is quote;
+      # let the leading - be optional on input, as sometimes omitted.
+      $a =~ s:^(.+)\[[^][]+\](|\-)$:$1-:; # e.g. AB[SOLUTE]- -> AB-;
+      # let the trailing - be optional on input, as sometimes omitted.
+      $a =~ s:([^][]+)\[.+\]$:$1:; # e.g. EX[SPECIALLY]-/ESPECIALLY] -> EX-
+      # which is a  mistake in the input.
+      $a =~ s:^\{(.+)\}$:$1:;                 # e.g. {YUPPIEDOM} -> YUPPIEDOM
+      $a =~ s:[A-Z]\[([^][])+\][A-Z]:$1-$3:i; # e.g. AMMU[N]IT- -> AMMU-IT-
+      $a =~ s:_\d$::;                         # e.g. THEM_1 -> THEM
+    }
+    $A[$n] = $a;
+  }
+  print join(" ", @A) . "\n";
+}
diff --git a/egs/swbd/ASR/local/swbd1_prepare_dict.sh b/egs/swbd/ASR/local/swbd1_prepare_dict.sh
new file mode 100755
index 000000000..eff5fb5f1
--- /dev/null
+++ b/egs/swbd/ASR/local/swbd1_prepare_dict.sh
@@ -0,0 +1,101 @@
+#!/usr/bin/env bash
+
+# Formatting the Mississippi State dictionary for use in Edinburgh. Differs
+# from the one in Kaldi s5 recipe in that it uses lower-case --Arnab (Jan 2013)
+
+# To be run from one directory above this script.
+
+#check existing directories
+[ $# != 0 ] && echo "Usage: local/swbd1_data_prep.sh" && exit 1
+
+srcdir=. # This is where we downloaded some stuff..
+dir=./data/local/dict_nosp
+mkdir -p $dir
+srcdict=$srcdir/swb_ms98_transcriptions/sw-ms98-dict.text
+
+# assume swbd_p1_data_prep.sh was done already.
+[ ! -f "$srcdict" ] && echo "$0: No such file $srcdict" && exit 1
+
+cp $srcdict $dir/lexicon0.txt || exit 1
+chmod a+w $dir/lexicon0.txt
+patch <local/dict.patch $dir/lexicon0.txt || exit 1
+
+#(2a) Dictionary preparation:
+# Pre-processing (remove comments)
+grep -v '^#' $dir/lexicon0.txt | awk 'NF>0' | sort >$dir/lexicon1.txt || exit 1
+
+cat $dir/lexicon1.txt | awk '{ for(n=2;n<=NF;n++){ phones[$n] = 1; }} END{for (p in phones) print p;}' |
+    grep -v sil >$dir/nonsilence_phones.txt || exit 1
+
+(
+    echo sil
+    echo spn
+    echo nsn
+    echo lau
+) >$dir/silence_phones.txt
+
+echo sil >$dir/optional_silence.txt
+
+# No "extra questions" in the input to this setup, as we don't
+# have stress or tone.
+echo -n >$dir/extra_questions.txt
+
+cp local/MSU_single_letter.txt $dir/
+# Add to the lexicon the silences, noises etc.
+# Add single letter lexicon
+# The original swbd lexicon does not have precise single letter lexicion
+# e.g. it does not have entry of W
+(
+    echo '!SIL SIL'
+    echo '[VOCALIZED-NOISE] spn'
+    echo '[NOISE] nsn'
+    echo '[LAUGHTER] lau'
+    echo '<UNK> spn'
+) |
+    cat - $dir/lexicon1.txt $dir/MSU_single_letter.txt >$dir/lexicon2.txt || exit 1
+
+# Map the words in the lexicon.  That is-- for each word in the lexicon, we map it
+# to a new written form.  The transformations we do are:
+# remove laughter markings, e.g.
+# [LAUGHTER-STORY] -> STORY
+# Remove partial-words, e.g.
+# -[40]1K W AH N K EY
+# becomes -1K
+# and
+# -[AN]Y IY
+# becomes
+# -Y
+# -[A]B[OUT]- B
+# becomes
+# -B-
+# Also, curly braces, which appear to be used for "nonstandard"
+# words or non-words, are removed, e.g.
+# {WOLMANIZED} W OW L M AX N AY Z D
+# -> WOLMANIZED
+# Also, mispronounced words, e.g.
+#  [YEAM/YEAH] Y AE M
+# are changed to just e.g. YEAM, i.e. the orthography
+# of the mispronounced version.
+# Note-- this is only really to be used in training.  The main practical
+# reason is to avoid having tons of disambiguation symbols, which
+# we otherwise would get because there are many partial words with
+# the same phone sequences (most problematic: S).
+# Also, map
+# THEM_1 EH M -> THEM
+# so that multiple pronunciations just have alternate entries
+# in the lexicon.
+
+local/swbd1_map_words.pl -f 1 $dir/lexicon2.txt | sort -u \
+    >$dir/lexicon3.txt || exit 1
+
+python3 local/format_acronyms_dict.py -i $dir/lexicon3.txt -o $dir/lexicon4.txt \
+    -L $dir/MSU_single_letter.txt -M $dir/acronyms_raw.map
+cat $dir/acronyms_raw.map | sort -u >$dir/acronyms.map
+
+(echo 'i ay') | cat - $dir/lexicon4.txt | tr '[A-Z]' '[a-z]' | sort -u >$dir/lexicon5.txt
+
+pushd $dir >&/dev/null
+ln -sf lexicon5.txt lexicon.txt # This is the final lexicon.
+popd >&/dev/null
+rm $dir/lexiconp.txt 2>/dev/null
+echo Prepared input dictionary and phone-sets for Switchboard phase 1.
diff --git a/egs/swbd/ASR/local/train_bpe_model.py b/egs/swbd/ASR/local/train_bpe_model.py
new file mode 100755
index 000000000..9b4e28635
--- /dev/null
+++ b/egs/swbd/ASR/local/train_bpe_model.py
@@ -0,0 +1,102 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# You can install sentencepiece via:
+#
+#  pip install sentencepiece
+#
+# Due to an issue reported in
+# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030
+#
+# Please install a version >=0.1.96
+
+import argparse
+import shutil
+from pathlib import Path
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        The generated bpe.model is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="Training transcript.",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        help="Vocabulary size for BPE training",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    vocab_size = args.vocab_size
+    lang_dir = Path(args.lang_dir)
+
+    model_type = "unigram"
+
+    model_prefix = f"{lang_dir}/{model_type}_{vocab_size}"
+    train_text = args.transcript
+    character_coverage = 1.0
+    input_sentence_size = 100000000
+
+    user_defined_symbols = ["<blk>", "<sos/eos>"]
+    unk_id = len(user_defined_symbols)
+    # Note: unk_id is fixed to 2.
+    # If you change it, you should also change other
+    # places that are using it.
+
+    user_defined_symbols += ["[LAUGHTER]", "[NOISE]", "[VOCALIZED-NOISE]"]
+
+    model_file = Path(model_prefix + ".model")
+    if not model_file.is_file():
+        spm.SentencePieceTrainer.train(
+            input=train_text,
+            vocab_size=vocab_size,
+            model_type=model_type,
+            model_prefix=model_prefix,
+            input_sentence_size=input_sentence_size,
+            character_coverage=character_coverage,
+            user_defined_symbols=user_defined_symbols,
+            unk_id=unk_id,
+            bos_id=-1,
+            eos_id=-1,
+        )
+    else:
+        print(f"{model_file} exists - skipping")
+        return
+
+    shutil.copyfile(model_file, f"{lang_dir}/bpe.model")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/swbd/ASR/local/validate_bpe_lexicon.py b/egs/swbd/ASR/local/validate_bpe_lexicon.py
new file mode 120000
index 000000000..721bb48e7
--- /dev/null
+++ b/egs/swbd/ASR/local/validate_bpe_lexicon.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/validate_bpe_lexicon.py
\ No newline at end of file
diff --git a/egs/swbd/ASR/prepare.sh b/egs/swbd/ASR/prepare.sh
new file mode 100755
index 000000000..47d12613b
--- /dev/null
+++ b/egs/swbd/ASR/prepare.sh
@@ -0,0 +1,463 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. Most of them can't be downloaded automatically
+# as they are not publically available and require a license purchased
+# from the LDC.
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=./download
+# swbd1_dir="/export/corpora3/LDC/LDC97S62"
+swbd1_dir=./download/LDC97S62/
+
+# eval2000_dir contains the following files and directories
+# downloaded from LDC website:
+#  - LDC2002S09
+#       - hub5e_00
+#  - LDC2002T43
+#       - reference
+eval2000_dir="/export/corpora2/LDC/eval2000"
+
+rt03_dir="/export/corpora/LDC/LDC2007S10"
+fisher_dir="/export/corpora3/LDC/LDC2004T19"
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+    # 5000
+    # 2000
+    1000
+    500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+    # This function is from espnet
+    local fname=${BASH_SOURCE[1]##*/}
+    echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "swbd1_dir: $swbd1_dir"
+log "eval2000_dir: $eval2000_dir"
+log "rt03_dir: $rt03_dir"
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+    log "Stage 1: Prepare SwitchBoard manifest"
+    # We assume that you have downloaded the SwitchBoard corpus
+    # to respective dirs
+    mkdir -p data/manifests
+    if [ ! -e data/manifests/.swbd.done ]; then
+        lhotse prepare switchboard --absolute-paths 1 --omit-silence $swbd1_dir data/manifests/swbd
+        ./local/normalize_and_filter_supervisions.py \
+            data/manifests/swbd/swbd_supervisions_all.jsonl.gz \
+            data/manifests/swbd/swbd_supervisions_all_norm.jsonl.gz
+        mv data/manifests/swbd/swbd_supervisions_all.jsonl.gz data/manifests/swbd/swbd_supervisions_orig.jsonl.gz
+        mv data/manifests/swbd/swbd_supervisions_all_norm.jsonl.gz data/manifests/swbd/swbd_supervisions_all.jsonl.gz
+
+        lhotse cut simple \
+            -r data/manifests/swbd/swbd_recordings_all.jsonl.gz \
+            -s data/manifests/swbd/swbd_supervisions_all.jsonl.gz \
+            data/manifests/swbd/swbd_train_all.jsonl.gz
+        lhotse cut trim-to-supervisions \
+            --discard-overlapping \
+            --discard-extra-channels \
+            data/manifests/swbd/swbd_train_all.jsonl.gz \
+            data/manifests/swbd/swbd_train_all_trimmed.jsonl.gz
+
+        num_splits=16
+        mkdir -p data/manifests/swbd_split${num_splits}
+        lhotse split ${num_splits} \
+            data/manifests/swbd/swbd_train_all_trimmed.jsonl.gz \
+            data/manifests/swbd_split${num_splits}
+
+        lhotse prepare eval2000 --absolute-paths 1 $eval2000_dir data/manifests/eval2000
+        ./local/normalize_eval2000.py \
+            data/manifests/eval2000/eval2000_supervisions_unnorm.jsonl.gz \
+            data/manifests/eval2000/eval2000_supervisions_all.jsonl.gz
+        
+        lhotse cut simple \
+            -r data/manifests/eval2000/eval2000_recordings_all.jsonl.gz \
+            -s data/manifests/eval2000/eval2000_supervisions_all.jsonl.gz \
+            data/manifests/eval2000/eval2000_cuts_all.jsonl.gz
+
+        lhotse cut trim-to-supervisions \
+            --discard-overlapping \
+            --discard-extra-channels \
+            data/manifests/eval2000/eval2000_cuts_all.jsonl.gz \
+            data/manifests/eval2000/eval2000_cuts_all_trimmed.jsonl.gz
+
+        sed -e 's:((:(:' -e 's:<B_ASIDE>::g' -e 's:<E_ASIDE>::g' \
+            $eval2000_dir/LDC2002T43/reference/hub5e00.english.000405.stm >  data/manifests/eval2000/stm
+        cp $eval2000_dir/LDC2002T43/reference/en20000405_hub5.glm  $dir/glm
+
+        # ./local/rt03_data_prep.sh $rt03_dir
+
+        # normalize eval2000 and rt03 texts by
+        # 1) convert upper to lower
+        # 2) remove tags (%AH) (%HESITATION) (%UH)
+        # 3) remove <B_ASIDE> <E_ASIDE>
+        # 4) remove "(" or ")"
+        # for x in  rt03; do
+        #     cp data/local/${x}/text data/local/${x}/text.org
+        #     paste -d "" \
+        #         <(cut -f 1 -d" " data/local/${x}/text.org) \
+        #         <(awk '{$1=""; print tolower($0)}' data/local/${x}/text.org | perl -pe 's| \(\%.*\)||g' | perl -pe 's| \<.*\>||g' | sed -e "s/(//g" -e "s/)//g") |
+        #         sed -e 's/\s\+/ /g' >data/local/${x}/text
+        #     rm data/local/${x}/text.org
+        # done
+
+        # lhotse fix data/manifests_rt03/swbd_recordings_rt03.jsonl.gz data/manifests_rt03/swbd_supervisions_rt03.jsonl.gz data/manifests
+
+        touch data/manifests/.swbd.done
+    fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+    log "Stage 2: Prepare musan manifest"
+    # We assume that you have downloaded the musan corpus
+    # to $dl_dir/musan
+    mkdir -p data/manifests
+    if [ ! -e data/manifests/.musan.done ]; then
+        lhotse prepare musan $dl_dir/musan data/manifests
+        touch data/manifests/.musan.done
+    fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+    log "Stage 3 I: Compute fbank for SwitchBoard"
+    if [ ! -e data/fbank/.swbd.done ]; then
+        mkdir -p data/fbank/swbd_split${num_splits}/
+        for index in $(seq 1 16); do
+            ./local/compute_fbank_swbd.py --split-index ${index} &
+        done
+        wait
+        pieces=$(find data/fbank/swbd_split${num_splits} -name "swbd_cuts_all.*.jsonl.gz")
+        lhotse combine $pieces data/fbank/swbd_cuts_all.jsonl.gz
+        touch data/fbank/.swbd.done
+    fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+    log "Stage 3 II: Compute fbank for eval2000"
+    if [ ! -e data/fbank/.eval2000.done ]; then
+        mkdir -p data/fbank/eval2000/
+        ./local/compute_fbank_eval2000.py 
+        touch data/fbank/.eval2000.done
+    fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+    log "Stage 4: Compute fbank for musan"
+    mkdir -p data/fbank
+    if [ ! -e data/fbank/.musan.done ]; then
+        ./local/compute_fbank_musan.py
+        touch data/fbank/.musan.done
+    fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+    log "Stage 5: Prepare phone based lang"
+    lang_dir=data/lang_phone
+    mkdir -p $lang_dir
+
+    if ! which jq; then
+      echo "This script is intended to be used with jq but you have not installed jq
+      Note: in Linux, you can install jq with the following command:
+      1. wget -O jq https://github.com/stedolan/jq/releases/download/jq-1.6/jq-linux64
+      2. chmod +x ./jq
+      3. cp jq /usr/bin" && exit 1
+    fi
+    if [ ! -f $lang_dir/text ] || [ ! -s $lang_dir/text ]; then
+        log "Prepare text."
+        gunzip -c data/manifests/swbd/swbd_supervisions_all.jsonl.gz \
+        | jq '.text' | sed 's/"//g'  > $lang_dir/text
+    fi
+
+    log "Prepare dict"
+    ./local/swbd1_prepare_dict.sh 
+    cut -f 2- -d" " $lang_dir/text >${lang_dir}/input.txt
+    # [noise] nsn
+    # !sil sil
+    # <unk> spn
+    cat data/local/dict_nosp/lexicon.txt | sed 's/-//g' | sed 's/\[vocalizednoise\]/\[vocalized-noise\]/g' |
+        sort | uniq >$lang_dir/lexicon_lower.txt
+
+    cat $lang_dir/lexicon_lower.txt | tr a-z A-Z > $lang_dir/lexicon.txt
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+        ./local/prepare_lang.py --lang-dir $lang_dir
+    fi
+
+    if [ ! -f $lang_dir/L.fst ]; then
+        log "Converting L.pt to L.fst"
+        ./shared/convert-k2-to-openfst.py \
+            --olabels aux_labels \
+            $lang_dir/L.pt \
+            $lang_dir/L.fst
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.fst ]; then
+        log "Converting L_disambig.pt to L_disambig.fst"
+        ./shared/convert-k2-to-openfst.py \
+            --olabels aux_labels \
+            $lang_dir/L_disambig.pt \
+            $lang_dir/L_disambig.fst
+    fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+    log "Stage 6: Prepare BPE based lang"
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        lang_dir=data/lang_bpe_${vocab_size}
+        mkdir -p $lang_dir
+        # We reuse words.txt from phone based lexicon
+        # so that the two can share G.pt later.
+        cp data/lang_phone/words.txt $lang_dir
+
+        if [ ! -f $lang_dir/transcript_words.txt ]; then
+            log "Generate data for BPE training"
+
+            cat data/lang_phone/text | cut -d " " -f 2- >$lang_dir/transcript_words.txt
+        fi
+
+        if [ ! -f $lang_dir/bpe.model ]; then
+            ./local/train_bpe_model.py \
+                --lang-dir $lang_dir \
+                --vocab-size $vocab_size \
+                --transcript $lang_dir/transcript_words.txt
+        fi
+
+        if [ ! -f $lang_dir/L_disambig.pt ]; then
+            ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+
+            log "Validating $lang_dir/lexicon.txt"
+            ./local/validate_bpe_lexicon.py \
+                --lexicon $lang_dir/lexicon.txt \
+                --bpe-model $lang_dir/bpe.model
+        fi
+
+        if [ ! -f $lang_dir/L.fst ]; then
+            log "Converting L.pt to L.fst"
+            ./shared/convert-k2-to-openfst.py \
+                --olabels aux_labels \
+                $lang_dir/L.pt \
+                $lang_dir/L.fst
+        fi
+
+        if [ ! -f $lang_dir/L_disambig.fst ]; then
+            log "Converting L_disambig.pt to L_disambig.fst"
+            ./shared/convert-k2-to-openfst.py \
+                --olabels aux_labels \
+                $lang_dir/L_disambig.pt \
+                $lang_dir/L_disambig.fst
+        fi
+    done
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+    log "Stage 7: Prepare bigram token-level P for MMI training"
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        lang_dir=data/lang_bpe_${vocab_size}
+
+        if [ ! -f $lang_dir/transcript_tokens.txt ]; then
+            ./local/convert_transcript_words_to_tokens.py \
+                --lexicon $lang_dir/lexicon.txt \
+                --transcript $lang_dir/transcript_words.txt \
+                --oov "<UNK>" \
+                >$lang_dir/transcript_tokens.txt
+        fi
+
+        if [ ! -f $lang_dir/P.arpa ]; then
+            ./shared/make_kn_lm.py \
+                -ngram-order 2 \
+                -text $lang_dir/transcript_tokens.txt \
+                -lm $lang_dir/P.arpa
+        fi
+
+        if [ ! -f $lang_dir/P.fst.txt ]; then
+            python3 -m kaldilm \
+                --read-symbol-table="$lang_dir/tokens.txt" \
+                --disambig-symbol='#0' \
+                --max-order=2 \
+                $lang_dir/P.arpa >$lang_dir/P.fst.txt
+        fi
+    done
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+    log "Stage 8: Prepare G"
+    lang_dir=data/lang_phone
+    # We assume you have install kaldilm, if not, please install
+    # it using: pip install kaldilm
+
+    mkdir -p data/lm
+    if [ ! -f data/lm/G_3_gram.fst.txt ]; then
+        # It is used in building HLG
+        ./shared/make_kn_lm.py \
+            -ngram-order 3 \
+            -text ${lang_dir}/input.txt \
+            -lm data/lm/3-gram.arpa
+        python3 -m kaldilm \
+            --read-symbol-table="data/lang_phone/words.txt" \
+            --disambig-symbol='#0' \
+            --max-order=3 \
+            data/lm/3-gram.arpa >data/lm/G_3_gram.fst.txt
+    fi
+
+    if [ ! -f data/lm/G_4_gram.fst.txt ]; then
+        # It is used for LM rescoring
+        ./shared/make_kn_lm.py \
+            -ngram-order 4 \
+            -text ${lang_dir}/input.txt \
+            -lm data/lm/4-gram.arpa
+        python3 -m kaldilm \
+            --read-symbol-table="data/lang_phone/words.txt" \
+            --disambig-symbol='#0' \
+            --max-order=4 \
+            data/lm/4-gram.arpa >data/lm/G_4_gram.fst.txt
+    fi
+fi
+
+if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
+    log "Stage 9: Compile HLG"
+    ./local/compile_hlg.py --lang-dir data/lang_phone
+
+    # Note If ./local/compile_hlg.py throws OOM,
+    # please switch to the following command
+    #
+    # ./local/compile_hlg_using_openfst.py --lang-dir data/lang_phone
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        lang_dir=data/lang_bpe_${vocab_size}
+        ./local/compile_hlg.py --lang-dir $lang_dir
+
+        # Note If ./local/compile_hlg.py throws OOM,
+        # please switch to the following command
+        #
+        # ./local/compile_hlg_using_openfst.py --lang-dir $lang_dir
+    done
+fi
+
+# Compile LG for RNN-T fast_beam_search decoding
+if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
+    log "Stage 10: Compile LG"
+    ./local/compile_lg.py --lang-dir data/lang_phone
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        lang_dir=data/lang_bpe_${vocab_size}
+        ./local/compile_lg.py --lang-dir $lang_dir
+    done
+fi
+
+if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
+    log "Stage 11: Generate LM training data"
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        log "Processing vocab_size == ${vocab_size}"
+        lang_dir=data/lang_bpe_${vocab_size}
+        out_dir=data/lm_training_bpe_${vocab_size}
+        mkdir -p $out_dir
+
+        if [ ! -f $out_dir/train.txt ]; then
+              tail -n 250000 data/lang_phone/input.txt > $out_dir/train.txt
+        fi
+
+        ./local/prepare_lm_training_data.py \
+            --bpe-model $lang_dir/bpe.model \
+            --lm-data data/lang_phone/input.txt \
+            --lm-archive $out_dir/lm_data.pt
+    done
+fi
+
+if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then
+    log "Stage 12: Generate LM validation data"
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        log "Processing vocab_size == ${vocab_size}"
+        out_dir=data/lm_training_bpe_${vocab_size}
+        mkdir -p $out_dir
+
+        if [ ! -f $out_dir/valid.txt ]; then
+            head -n 14332 data/lang_phone/input.txt > $out_dir/valid.txt
+        fi
+
+        lang_dir=data/lang_bpe_${vocab_size}
+        ./local/prepare_lm_training_data.py \
+            --bpe-model $lang_dir/bpe.model \
+            --lm-data $out_dir/valid.txt \
+            --lm-archive $out_dir/lm_data-valid.pt
+    done
+fi
+
+if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then
+    log "Stage 13: Generate LM test data"
+    testsets=(eval2000)
+
+    for testset in ${testsets[@]}; do
+        for vocab_size in ${vocab_sizes[@]}; do
+            log "Processing vocab_size == ${vocab_size}"
+            out_dir=data/lm_training_bpe_${vocab_size}
+            mkdir -p $out_dir
+
+            if [ ! -f $out_dir/${testset}.txt ]; then
+                gunzip -c data/manifests/${testset}/eval2000_supervisions_all.jsonl.gz \
+                    | jq '.text' | sed 's/"//g' > $out_dir/${testset}.txt
+            fi
+
+            lang_dir=data/lang_bpe_${vocab_size}
+            ./local/prepare_lm_training_data.py \
+                --bpe-model $lang_dir/bpe.model \
+                --lm-data $out_dir/${testset}.txt \
+                --lm-archive $out_dir/lm_data-${testset}.pt
+        done
+    done
+fi
+
+if [ $stage -le 14 ] && [ $stop_stage -ge 14 ]; then
+    log "Stage 14: Sort LM training data"
+    testsets=(eval2000)
+    # Sort LM training data by sentence length in descending order
+    # for ease of training.
+    #
+    # Sentence length equals to the number of BPE tokens
+    # in a sentence.
+
+    for vocab_size in ${vocab_sizes[@]}; do
+        out_dir=data/lm_training_bpe_${vocab_size}
+        mkdir -p $out_dir
+        ./local/sort_lm_training_data.py \
+            --in-lm-data $out_dir/lm_data.pt \
+            --out-lm-data $out_dir/sorted_lm_data.pt \
+            --out-statistics $out_dir/statistics.txt
+        for testset in ${testsets[@]}; do
+            ./local/sort_lm_training_data.py \
+                --in-lm-data $out_dir/lm_data-${testset}.pt \
+                --out-lm-data $out_dir/sorted_lm_data-${testset}.pt \
+                --out-statistics $out_dir/statistics-test-${testset}.txt
+        done
+    done
+fi
diff --git a/egs/swbd/ASR/shared b/egs/swbd/ASR/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/swbd/ASR/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/egs/swbd/ASR/utils/filter_scp.pl b/egs/swbd/ASR/utils/filter_scp.pl
new file mode 100755
index 000000000..b76d37f41
--- /dev/null
+++ b/egs/swbd/ASR/utils/filter_scp.pl
@@ -0,0 +1,87 @@
+#!/usr/bin/env perl
+# Copyright 2010-2012 Microsoft Corporation
+#                     Johns Hopkins University (author: Daniel Povey)
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+
+# This script takes a list of utterance-ids or any file whose first field
+# of each line is an utterance-id, and filters an scp
+# file (or any file whose "n-th" field is an utterance id), printing
+# out only those lines whose "n-th" field is in id_list. The index of
+# the "n-th" field is 1, by default, but can be changed by using
+# the -f <n> switch
+
+$exclude = 0;
+$field = 1;
+$shifted = 0;
+
+do {
+  $shifted=0;
+  if ($ARGV[0] eq "--exclude") {
+    $exclude = 1;
+    shift @ARGV;
+    $shifted=1;
+  }
+  if ($ARGV[0] eq "-f") {
+    $field = $ARGV[1];
+    shift @ARGV; shift @ARGV;
+    $shifted=1
+  }
+} while ($shifted);
+
+if(@ARGV < 1 || @ARGV > 2) {
+  die "Usage: filter_scp.pl [--exclude] [-f <field-to-filter-on>] id_list [in.scp] > out.scp \n" .
+      "Prints only the input lines whose f'th field (default: first) is in 'id_list'.\n" .
+      "Note: only the first field of each line in id_list matters.  With --exclude, prints\n" .
+      "only the lines that were *not* in id_list.\n" .
+      "Caution: previously, the -f option was interpreted as a zero-based field index.\n" .
+      "If your older scripts (written before Oct 2014) stopped working and you used the\n" .
+      "-f option, add 1 to the argument.\n" .
+      "See also: utils/filter_scp.pl .\n";
+}
+
+
+$idlist = shift @ARGV;
+open(F, "<$idlist") || die "Could not open id-list file $idlist";
+while(<F>) {
+  @A = split;
+  @A>=1 || die "Invalid id-list file line $_";
+  $seen{$A[0]} = 1;
+}
+
+if ($field == 1) { # Treat this as special case, since it is common.
+  while(<>) {
+    $_ =~ m/\s*(\S+)\s*/ || die "Bad line $_, could not get first field.";
+    # $1 is what we filter on.
+    if ((!$exclude && $seen{$1}) || ($exclude && !defined $seen{$1})) {
+      print $_;
+    }
+  }
+} else {
+  while(<>) {
+    @A = split;
+    @A > 0 || die "Invalid scp file line $_";
+    @A >= $field || die "Invalid scp file line $_";
+    if ((!$exclude && $seen{$A[$field-1]}) || ($exclude && !defined $seen{$A[$field-1]})) {
+      print $_;
+    }
+  }
+}
+
+# tests:
+# the following should print "foo 1"
+# ( echo foo 1; echo bar 2 ) | utils/filter_scp.pl <(echo foo)
+# the following should print "bar 2".
+# ( echo foo 1; echo bar 2 ) | utils/filter_scp.pl -f 2 <(echo 2)
diff --git a/egs/swbd/ASR/utils/fix_data_dir.sh b/egs/swbd/ASR/utils/fix_data_dir.sh
new file mode 100755
index 000000000..ca0972ca8
--- /dev/null
+++ b/egs/swbd/ASR/utils/fix_data_dir.sh
@@ -0,0 +1,197 @@
+#!/bin/bash
+
+# This script makes sure that only the segments present in
+# all of "feats.scp", "wav.scp" [if present], segments [if present]
+# text, and utt2spk are present in any of them.
+# It puts the original contents of data-dir into
+# data-dir/.backup
+
+cmd="$@"
+
+utt_extra_files=
+spk_extra_files=
+
+. utils/parse_options.sh
+
+if [ $# != 1 ]; then
+  echo "Usage: utils/data/fix_data_dir.sh <data-dir>"
+  echo "e.g.: utils/data/fix_data_dir.sh data/train"
+  echo "This script helps ensure that the various files in a data directory"
+  echo "are correctly sorted and filtered, for example removing utterances"
+  echo "that have no features (if feats.scp is present)"
+  exit 1
+fi
+
+data=$1
+
+if [ -f $data/images.scp ]; then
+  image/fix_data_dir.sh $cmd
+  exit $?
+fi
+
+mkdir -p $data/.backup
+
+[ ! -d $data ] && echo "$0: no such directory $data" && exit 1;
+
+[ ! -f $data/utt2spk ] && echo "$0: no such file $data/utt2spk" && exit 1;
+
+set -e -o pipefail -u
+
+tmpdir=$(mktemp -d /tmp/kaldi.XXXX);
+trap 'rm -rf "$tmpdir"' EXIT HUP INT PIPE TERM
+
+export LC_ALL=C
+
+function check_sorted {
+  file=$1
+  sort -k1,1 -u <$file >$file.tmp
+  if ! cmp -s $file $file.tmp; then
+    echo "$0: file $1 is not in sorted order or not unique, sorting it"
+    mv $file.tmp $file
+  else
+    rm $file.tmp
+  fi
+}
+
+for x in utt2spk spk2utt feats.scp text segments wav.scp cmvn.scp vad.scp \
+    reco2file_and_channel spk2gender utt2lang utt2uniq utt2dur reco2dur utt2num_frames; do
+  if [ -f $data/$x ]; then
+    cp $data/$x $data/.backup/$x
+    check_sorted $data/$x
+  fi
+done
+
+
+function filter_file {
+  filter=$1
+  file_to_filter=$2
+  cp $file_to_filter ${file_to_filter}.tmp
+  utils/filter_scp.pl $filter ${file_to_filter}.tmp > $file_to_filter
+  if ! cmp ${file_to_filter}.tmp  $file_to_filter >&/dev/null; then
+    length1=$(cat ${file_to_filter}.tmp | wc -l)
+    length2=$(cat ${file_to_filter} | wc -l)
+    if [ $length1 -ne $length2 ]; then
+      echo "$0: filtered $file_to_filter from $length1 to $length2 lines based on filter $filter."
+    fi
+  fi
+  rm $file_to_filter.tmp
+}
+
+function filter_recordings {
+  # We call this once before the stage when we filter on utterance-id, and once
+  # after.
+
+  if [ -f $data/segments ]; then
+  # We have a segments file -> we need to filter this and the file wav.scp, and
+  # reco2file_and_utt, if it exists, to make sure they have the same list of
+  # recording-ids.
+
+    if [ ! -f $data/wav.scp ]; then
+      echo "$0: $data/segments exists but not $data/wav.scp"
+      exit 1;
+    fi
+    awk '{print $2}' < $data/segments | sort | uniq > $tmpdir/recordings
+    n1=$(cat $tmpdir/recordings | wc -l)
+    [ ! -s $tmpdir/recordings ] && \
+      echo "Empty list of recordings (bad file $data/segments)?" && exit 1;
+    utils/filter_scp.pl $data/wav.scp $tmpdir/recordings > $tmpdir/recordings.tmp
+    mv $tmpdir/recordings.tmp $tmpdir/recordings
+
+
+    cp $data/segments{,.tmp}; awk '{print $2, $1, $3, $4}' <$data/segments.tmp >$data/segments
+    filter_file $tmpdir/recordings $data/segments
+    cp $data/segments{,.tmp}; awk '{print $2, $1, $3, $4}' <$data/segments.tmp >$data/segments
+    rm $data/segments.tmp
+
+    filter_file $tmpdir/recordings $data/wav.scp
+    [ -f $data/reco2file_and_channel ] && filter_file $tmpdir/recordings $data/reco2file_and_channel
+    [ -f $data/reco2dur ] && filter_file $tmpdir/recordings $data/reco2dur
+    true
+  fi
+}
+
+function filter_speakers {
+  # throughout this program, we regard utt2spk as primary and spk2utt as derived, so...
+  utils/utt2spk_to_spk2utt.pl $data/utt2spk > $data/spk2utt
+
+  cat $data/spk2utt | awk '{print $1}' > $tmpdir/speakers
+  for s in cmvn.scp spk2gender; do
+    f=$data/$s
+    if [ -f $f ]; then
+      filter_file $f $tmpdir/speakers
+    fi
+  done
+
+  filter_file $tmpdir/speakers $data/spk2utt
+  utils/spk2utt_to_utt2spk.pl $data/spk2utt > $data/utt2spk
+
+  for s in cmvn.scp spk2gender $spk_extra_files; do
+    f=$data/$s
+    if [ -f $f ]; then
+      filter_file $tmpdir/speakers $f
+    fi
+  done
+}
+
+function filter_utts {
+  cat $data/utt2spk | awk '{print $1}' > $tmpdir/utts
+
+  ! cat $data/utt2spk | sort | cmp - $data/utt2spk && \
+    echo "utt2spk is not in sorted order (fix this yourself)" && exit 1;
+
+  ! cat $data/utt2spk | sort -k2 | cmp - $data/utt2spk && \
+    echo "utt2spk is not in sorted order when sorted first on speaker-id " && \
+    echo "(fix this by making speaker-ids prefixes of utt-ids)" && exit 1;
+
+  ! cat $data/spk2utt | sort | cmp - $data/spk2utt && \
+    echo "spk2utt is not in sorted order (fix this yourself)" && exit 1;
+
+  if [ -f $data/utt2uniq ]; then
+    ! cat $data/utt2uniq | sort | cmp - $data/utt2uniq && \
+      echo "utt2uniq is not in sorted order (fix this yourself)" && exit 1;
+  fi
+
+  maybe_wav=
+  maybe_reco2dur=
+  [ ! -f $data/segments ] && maybe_wav=wav.scp # wav indexed by utts only if segments does not exist.
+  [ -s $data/reco2dur ] && [ ! -f $data/segments ] && maybe_reco2dur=reco2dur # reco2dur indexed by utts
+  for x in feats.scp text segments utt2lang $maybe_wav; do
+    if [ -f $data/$x ]; then
+      utils/filter_scp.pl $data/$x $tmpdir/utts > $tmpdir/utts.tmp
+      mv $tmpdir/utts.tmp $tmpdir/utts
+    fi
+  done
+  [ ! -s $tmpdir/utts ] && echo "fix_data_dir.sh: no utterances remained: not proceeding further." && \
+    rm $tmpdir/utts && exit 1;
+
+
+  if [ -f $data/utt2spk ]; then
+    new_nutts=$(cat $tmpdir/utts | wc -l)
+    old_nutts=$(cat $data/utt2spk | wc -l)
+    if [ $new_nutts -ne $old_nutts ]; then
+      echo "fix_data_dir.sh: kept $new_nutts utterances out of $old_nutts"
+    else
+      echo "fix_data_dir.sh: kept all $old_nutts utterances."
+    fi
+  fi
+
+  for x in utt2spk utt2uniq feats.scp vad.scp text segments utt2lang utt2dur utt2num_frames $maybe_wav $maybe_reco2dur $utt_extra_files; do
+    if [ -f $data/$x ]; then
+      cp $data/$x $data/.backup/$x
+      if ! cmp -s $data/$x <( utils/filter_scp.pl $tmpdir/utts $data/$x ) ; then
+        utils/filter_scp.pl $tmpdir/utts $data/.backup/$x > $data/$x
+      fi
+    fi
+  done
+
+}
+
+filter_recordings
+filter_speakers
+filter_utts
+filter_speakers
+filter_recordings
+
+utils/utt2spk_to_spk2utt.pl $data/utt2spk > $data/spk2utt
+
+echo "fix_data_dir.sh: old files are kept in $data/.backup"
diff --git a/egs/swbd/ASR/utils/parse_options.sh b/egs/swbd/ASR/utils/parse_options.sh
new file mode 100755
index 000000000..34476fdb3
--- /dev/null
+++ b/egs/swbd/ASR/utils/parse_options.sh
@@ -0,0 +1,97 @@
+#!/bin/bash
+
+# Copyright 2012  Johns Hopkins University (Author: Daniel Povey);
+#                 Arnab Ghoshal, Karel Vesely
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Parse command-line options.
+# To be sourced by another script (as in ". parse_options.sh").
+# Option format is: --option-name arg
+# and shell variable "option_name" gets set to value "arg."
+# The exception is --help, which takes no arguments, but prints the
+# $help_message variable (if defined).
+
+
+###
+### The --config file options have lower priority to command line
+### options, so we need to import them first...
+###
+
+# Now import all the configs specified by command-line, in left-to-right order
+for ((argpos=1; argpos<$#; argpos++)); do
+  if [ "${!argpos}" == "--config" ]; then
+    argpos_plus1=$((argpos+1))
+    config=${!argpos_plus1}
+    [ ! -r $config ] && echo "$0: missing config '$config'" && exit 1
+    . $config  # source the config file.
+  fi
+done
+
+
+###
+### No we process the command line options
+###
+while true; do
+  [ -z "${1:-}" ] && break;  # break if there are no arguments
+  case "$1" in
+    # If the enclosing script is called with --help option, print the help
+    # message and exit.  Scripts should put help messages in $help_message
+    --help|-h) if [ -z "$help_message" ]; then echo "No help found." 1>&2;
+      else printf "$help_message\n" 1>&2 ; fi;
+      exit 0 ;;
+    --*=*) echo "$0: options to scripts must be of the form --name value, got '$1'"
+      exit 1 ;;
+    # If the first command-line argument begins with "--" (e.g. --foo-bar),
+    # then work out the variable name as $name, which will equal "foo_bar".
+    --*) name=`echo "$1" | sed s/^--// | sed s/-/_/g`;
+      # Next we test whether the variable in question is undefned-- if so it's
+      # an invalid option and we die.  Note: $0 evaluates to the name of the
+      # enclosing script.
+      # The test [ -z ${foo_bar+xxx} ] will return true if the variable foo_bar
+      # is undefined.  We then have to wrap this test inside "eval" because
+      # foo_bar is itself inside a variable ($name).
+      eval '[ -z "${'$name'+xxx}" ]' && echo "$0: invalid option $1" 1>&2 && exit 1;
+
+      oldval="`eval echo \\$$name`";
+      # Work out whether we seem to be expecting a Boolean argument.
+      if [ "$oldval" == "true" ] || [ "$oldval" == "false" ]; then
+        was_bool=true;
+      else
+        was_bool=false;
+      fi
+
+      # Set the variable to the right value-- the escaped quotes make it work if
+      # the option had spaces, like --cmd "queue.pl -sync y"
+      eval $name=\"$2\";
+
+      # Check that Boolean-valued arguments are really Boolean.
+      if $was_bool && [[ "$2" != "true" && "$2" != "false" ]]; then
+        echo "$0: expected \"true\" or \"false\": $1 $2" 1>&2
+        exit 1;
+      fi
+      shift 2;
+      ;;
+  *) break;
+  esac
+done
+
+
+# Check for an empty argument to the --cmd option, which can easily occur as a
+# result of scripting errors.
+[ ! -z "${cmd+xxx}" ] && [ -z "$cmd" ] && echo "$0: empty argument to --cmd option" 1>&2 && exit 1;
+
+
+true; # so this script returns exit code 0.
diff --git a/egs/swbd/ASR/utils/spk2utt_to_utt2spk.pl b/egs/swbd/ASR/utils/spk2utt_to_utt2spk.pl
new file mode 100755
index 000000000..23992f25d
--- /dev/null
+++ b/egs/swbd/ASR/utils/spk2utt_to_utt2spk.pl
@@ -0,0 +1,27 @@
+#!/usr/bin/env perl
+# Copyright 2010-2011 Microsoft Corporation
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+
+while(<>){ 
+    @A = split(" ", $_);
+    @A > 1 || die "Invalid line in spk2utt file: $_";
+    $s = shift @A;
+    foreach $u ( @A ) {
+        print "$u $s\n";
+    }
+}
+
+
diff --git a/egs/swbd/ASR/utils/utt2spk_to_spk2utt.pl b/egs/swbd/ASR/utils/utt2spk_to_spk2utt.pl
new file mode 100755
index 000000000..6e0e438ca
--- /dev/null
+++ b/egs/swbd/ASR/utils/utt2spk_to_spk2utt.pl
@@ -0,0 +1,38 @@
+#!/usr/bin/env perl
+# Copyright 2010-2011 Microsoft Corporation
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#  http://www.apache.org/licenses/LICENSE-2.0
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+# converts an utt2spk file to a spk2utt file.
+# Takes input from the stdin or from a file argument;
+# output goes to the standard out.
+
+if ( @ARGV > 1 ) {
+    die "Usage: utt2spk_to_spk2utt.pl [ utt2spk ] > spk2utt";
+}
+
+while(<>){ 
+    @A = split(" ", $_);
+    @A == 2 || die "Invalid line in utt2spk file: $_";
+    ($u,$s) = @A;
+    if(!$seen_spk{$s}) {
+        $seen_spk{$s} = 1;
+        push @spklist, $s;
+    }
+    push (@{$spk_hash{$s}}, "$u");
+}
+foreach $s (@spklist) {
+    $l = join(' ',@{$spk_hash{$s}});
+    print "$s $l\n";
+}
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py
index 39beffdcf..5269a1778 100644
--- a/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py
@@ -228,7 +228,7 @@ class TAL_CSASRAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/train.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/train.py
index 417515968..d03970265 100755
--- a/egs/tal_csasr/ASR/pruned_transducer_stateless5/train.py
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/train.py
@@ -915,7 +915,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/train.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/train.py
index d80e0147c..aee3972cd 100755
--- a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/train.py
+++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/train.py
@@ -69,7 +69,7 @@ from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.utils.tensorboard import SummaryWriter
 from zipformer import Zipformer
 
-from icefall import diagnostics, byte_encode, tokenize_by_CJK_char
+from icefall import byte_encode, diagnostics, tokenize_by_CJK_char
 from icefall.checkpoint import load_checkpoint, remove_checkpoints
 from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
 from icefall.checkpoint import (
@@ -1018,7 +1018,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/tedlium3/ASR/conformer_ctc2/local b/egs/tedlium3/ASR/conformer_ctc2/local
new file mode 120000
index 000000000..c820590c5
--- /dev/null
+++ b/egs/tedlium3/ASR/conformer_ctc2/local
@@ -0,0 +1 @@
+../local
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/conformer_ctc2/train.py b/egs/tedlium3/ASR/conformer_ctc2/train.py
index 42e4c010a..fc3e3b2d9 100755
--- a/egs/tedlium3/ASR/conformer_ctc2/train.py
+++ b/egs/tedlium3/ASR/conformer_ctc2/train.py
@@ -905,7 +905,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/local b/egs/tedlium3/ASR/pruned_transducer_stateless/local
new file mode 120000
index 000000000..c820590c5
--- /dev/null
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/local
@@ -0,0 +1 @@
+../local
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/transducer_stateless/asr_datamodule.py b/egs/tedlium3/ASR/transducer_stateless/asr_datamodule.py
index 28d0d3826..d4a9e4bc9 100644
--- a/egs/tedlium3/ASR/transducer_stateless/asr_datamodule.py
+++ b/egs/tedlium3/ASR/transducer_stateless/asr_datamodule.py
@@ -205,7 +205,7 @@ class TedLiumAsrDataModule:
             logging.info("Enable MUSAN")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/tedlium3/ASR/transducer_stateless/local b/egs/tedlium3/ASR/transducer_stateless/local
new file mode 120000
index 000000000..c820590c5
--- /dev/null
+++ b/egs/tedlium3/ASR/transducer_stateless/local
@@ -0,0 +1 @@
+../local
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/local b/egs/tedlium3/ASR/zipformer/local
new file mode 120000
index 000000000..c820590c5
--- /dev/null
+++ b/egs/tedlium3/ASR/zipformer/local
@@ -0,0 +1 @@
+../local
\ No newline at end of file
diff --git a/egs/tedlium3/ASR/zipformer/train.py b/egs/tedlium3/ASR/zipformer/train.py
index 9271c8438..33d03908c 100755
--- a/egs/tedlium3/ASR/zipformer/train.py
+++ b/egs/tedlium3/ASR/zipformer/train.py
@@ -1126,7 +1126,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/timit/ASR/tdnn_lstm_ctc/asr_datamodule.py b/egs/timit/ASR/tdnn_lstm_ctc/asr_datamodule.py
index 7c299d601..5d1b3c367 100644
--- a/egs/timit/ASR/tdnn_lstm_ctc/asr_datamodule.py
+++ b/egs/timit/ASR/tdnn_lstm_ctc/asr_datamodule.py
@@ -157,7 +157,7 @@ class TimitAsrDataModule(DataModule):
         cuts_musan = load_manifest(self.args.feature_dir / "musan_cuts.jsonl.gz")
 
         logging.info("About to create train dataset")
-        transforms = [CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20))]
+        transforms = [CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20))]
         if self.args.concatenate_cuts:
             logging.info(
                 f"Using cut concatenation with duration factor "
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
index c5967f10a..1dbfb9709 100644
--- a/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
@@ -215,7 +215,7 @@ class WenetSpeechAsrDataModule:
         if self.args.enable_musan:
             logging.info("Enable MUSAN")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/decode.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/decode.py
index bdd1f27bc..2bafe25d6 100755
--- a/egs/wenetspeech/ASR/pruned_transducer_stateless2/decode.py
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/decode.py
@@ -203,7 +203,7 @@ def get_parser():
         "--beam-size",
         type=int,
         default=4,
-        help="""An interger indicating how many candidates we will keep for each
+        help="""An integer indicating how many candidates we will keep for each
         frame. Used only when --decoding-method is beam_search or
         modified_beam_search.""",
     )
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/finetune.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/finetune.py
index e703100a9..c34f1593d 100755
--- a/egs/wenetspeech/ASR/pruned_transducer_stateless2/finetune.py
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/finetune.py
@@ -78,7 +78,7 @@ def add_finetune_arguments(parser: argparse.ArgumentParser):
         default=None,
         help="""
         Modules to be initialized. It matches all parameters starting with
-        a specific key. The keys are given with Comma seperated. If None,
+        a specific key. The keys are given with Comma separated. If None,
         all modules will be initialised. For example, if you only want to
         initialise all parameters staring with "encoder", use "encoder";
         if you want to initialise parameters starting with encoder or decoder,
@@ -498,8 +498,12 @@ def load_model_params(
         dst_state_dict = model.state_dict()
         for module in init_modules:
             logging.info(f"Loading parameters starting with prefix {module}")
-            src_keys = [k for k in src_state_dict.keys() if k.startswith(module)]
-            dst_keys = [k for k in dst_state_dict.keys() if k.startswith(module)]
+            src_keys = [
+                k for k in src_state_dict.keys() if k.startswith(module.strip() + ".")
+            ]
+            dst_keys = [
+                k for k in dst_state_dict.keys() if k.startswith(module.strip() + ".")
+            ]
             assert set(src_keys) == set(dst_keys)  # two sets should match exactly
             for key in src_keys:
                 dst_state_dict[key] = src_state_dict.pop(key)
@@ -734,7 +738,7 @@ def train_one_epoch(
             scaler.update()
             optimizer.zero_grad()
         except:  # noqa
-            display_and_save_batch(batch, params=params)
+            display_and_save_batch(batch, params=params, graph_compiler=graph_compiler)
             raise
 
         if params.print_diagnostics and batch_idx == 5:
@@ -886,7 +890,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py
index 48b347b64..49977e01b 100644
--- a/egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py
@@ -851,7 +851,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/train.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/train.py
index 8e1b12dba..931e699d9 100755
--- a/egs/wenetspeech/ASR/pruned_transducer_stateless5/train.py
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/train.py
@@ -985,7 +985,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/wenetspeech/ASR/zipformer/train.py b/egs/wenetspeech/ASR/zipformer/train.py
index 83dbfa22f..b1557dedb 100755
--- a/egs/wenetspeech/ASR/zipformer/train.py
+++ b/egs/wenetspeech/ASR/zipformer/train.py
@@ -1128,7 +1128,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/asr_datamodule.py
index 6362ab7cd..7594fb28e 100644
--- a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/asr_datamodule.py
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/asr_datamodule.py
@@ -218,7 +218,7 @@ class Xbmu_AmdoAsrDataModule:
             logging.info("About to get Musan cuts")
             cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
             transforms.append(
-                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
             )
         else:
             logging.info("Disable MUSAN")
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/train.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/train.py
index 5b5ac17be..a6fa46b17 100755
--- a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/train.py
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/train.py
@@ -1001,7 +1001,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/train.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/train.py
index f8dd7b287..8c53972fd 100755
--- a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/train.py
+++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/train.py
@@ -993,7 +993,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2**22
+            512
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
 
diff --git a/egs/yesno/ASR/local/prepare_lang_fst.py b/egs/yesno/ASR/local/prepare_lang_fst.py
new file mode 120000
index 000000000..c5787c534
--- /dev/null
+++ b/egs/yesno/ASR/local/prepare_lang_fst.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_fst.py
\ No newline at end of file
diff --git a/egs/yesno/ASR/prepare.sh b/egs/yesno/ASR/prepare.sh
index d4ef8d601..41db0cf7c 100755
--- a/egs/yesno/ASR/prepare.sh
+++ b/egs/yesno/ASR/prepare.sh
@@ -60,6 +60,7 @@ if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
   ) > $lang_dir/lexicon.txt
 
   ./local/prepare_lang.py
+  ./local/prepare_lang_fst.py --lang-dir ./data/lang_phone --has-silence 1
 fi
 
 if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
diff --git a/egs/yesno/ASR/tdnn/jit_pretrained.py b/egs/yesno/ASR/tdnn/jit_pretrained.py
index 84390fca5..7581ecb83 100755
--- a/egs/yesno/ASR/tdnn/jit_pretrained.py
+++ b/egs/yesno/ASR/tdnn/jit_pretrained.py
@@ -156,7 +156,6 @@ def main():
 
     features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
 
-    # Note: We don't use key padding mask for attention during decoding
     nnet_output = model(features)
 
     batch_size = nnet_output.shape[0]
diff --git a/egs/yesno/ASR/tdnn/jit_pretrained_decode_with_H.py b/egs/yesno/ASR/tdnn/jit_pretrained_decode_with_H.py
new file mode 100755
index 000000000..ff8c742af
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/jit_pretrained_decode_with_H.py
@@ -0,0 +1,208 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This file shows how to use a torchscript model for decoding with H
+on CPU using OpenFST and decoders from kaldi.
+
+Usage:
+
+  ./tdnn/jit_pretrained_decode_with_H.py \
+    --nn-model ./tdnn/exp/cpu_jit.pt \
+    --H ./data/lang_phone/H.fst \
+    --tokens ./data/lang_phone/tokens.txt \
+    ./download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+    ./download/waves_yesno/0_0_1_0_0_0_1_0.wav \
+    ./download/waves_yesno/0_0_1_0_0_1_1_1.wav
+
+Note that to generate ./tdnn/exp/cpu_jit.pt,
+you can use ./export.py --jit 1
+"""
+
+import argparse
+import logging
+import math
+from typing import Dict, List
+
+import kaldifeat
+import kaldifst
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="""Path to the torchscript model.
+        You can use ./tdnn/export.py --jit 1
+        to obtain it
+        """,
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to tokens.txt",
+    )
+
+    parser.add_argument("--H", type=str, required=True, help="Path to H.fst")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def read_tokens(tokens_txt: str) -> Dict[int, str]:
+    id2token = dict()
+    with open(tokens_txt, encoding="utf-8") as f:
+        for line in f:
+            token, idx = line.strip().split()
+            id2token[int(idx)] = token
+
+    return id2token
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    nnet_output: torch.Tensor,
+    H: kaldifst,
+    id2token: Dict[int, str],
+) -> List[str]:
+    decodable = DecodableCtc(nnet_output)
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(H, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        print(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        print(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    # are shifted by 1 during graph construction
+    hyps = [id2token[i - 1] for i in osymbols_out if id2token[i - 1] != "SIL"]
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    logging.info("Loading torchscript model")
+    model = torch.jit.load(args.nn_model)
+    model.eval()
+    model.to(device)
+
+    logging.info(f"Loading H from {args.H}")
+    H = kaldifst.StdVectorFst.read(args.H)
+
+    sample_rate = 8000
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 23
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files, expected_sample_rate=sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    nnet_output = model(features)
+
+    id2token = read_tokens(args.tokens)
+
+    hyps = []
+    for i in range(nnet_output.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[0],
+            nnet_output=nnet_output[i],
+            H=H,
+            id2token=id2token,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/yesno/ASR/tdnn/jit_pretrained_decode_with_HL.py b/egs/yesno/ASR/tdnn/jit_pretrained_decode_with_HL.py
new file mode 100755
index 000000000..05ba74f9a
--- /dev/null
+++ b/egs/yesno/ASR/tdnn/jit_pretrained_decode_with_HL.py
@@ -0,0 +1,207 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This file shows how to use a torchscript model for decoding with HL
+on CPU using OpenFST and decoders from kaldi.
+
+Usage:
+
+  ./tdnn/jit_pretrained_decode_with_HL.py \
+    --nn-model ./tdnn/exp/cpu_jit.pt \
+    --HL ./data/lang_phone/HL.fst \
+    --words ./data/lang_phone/words.txt \
+    ./download/waves_yesno/0_0_0_1_0_0_0_1.wav \
+    ./download/waves_yesno/0_0_1_0_0_0_1_0.wav \
+    ./download/waves_yesno/0_0_1_0_0_1_1_1.wav
+
+Note that to generate ./tdnn/exp/cpu_jit.pt,
+you can use ./export.py --jit 1
+"""
+
+import argparse
+import logging
+import math
+from typing import Dict, List
+
+import kaldifeat
+import kaldifst
+import torch
+import torchaudio
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from torch.nn.utils.rnn import pad_sequence
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--nn-model",
+        type=str,
+        required=True,
+        help="""Path to the torchscript model.
+        You can use ./tdnn/export.py --jit 1
+        to obtain it
+        """,
+    )
+
+    parser.add_argument(
+        "--words",
+        type=str,
+        required=True,
+        help="Path to words.txt",
+    )
+
+    parser.add_argument("--HL", type=str, required=True, help="Path to HL.fst")
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. ",
+    )
+
+    return parser
+
+
+def read_words(words_txt: str) -> Dict[int, str]:
+    id2word = dict()
+    with open(words_txt, encoding="utf-8") as f:
+        for line in f:
+            word, idx = line.strip().split()
+            id2word[int(idx)] = word
+
+    return id2word
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        if sample_rate != expected_sample_rate:
+            wave = torchaudio.functional.resample(
+                wave,
+                orig_freq=sample_rate,
+                new_freq=expected_sample_rate,
+            )
+
+        # We use only the first channel
+        ans.append(wave[0].contiguous())
+    return ans
+
+
+def decode(
+    filename: str,
+    nnet_output: torch.Tensor,
+    HL: kaldifst,
+    id2word: Dict[int, str],
+) -> List[str]:
+    decodable = DecodableCtc(nnet_output)
+    decoder_opts = FasterDecoderOptions(max_active=3000)
+    decoder = FasterDecoder(HL, decoder_opts)
+    decoder.decode(decodable)
+
+    if not decoder.reached_final():
+        print(f"failed to decode {filename}")
+        return [""]
+
+    ok, best_path = decoder.get_best_path()
+
+    (
+        ok,
+        isymbols_out,
+        osymbols_out,
+        total_weight,
+    ) = kaldifst.get_linear_symbol_sequence(best_path)
+    if not ok:
+        print(f"failed to get linear symbol sequence for {filename}")
+        return [""]
+
+    hyps = [id2word[i] for i in osymbols_out if id2word[i] != "<SIL>"]
+
+    return hyps
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    device = torch.device("cpu")
+
+    logging.info(f"device: {device}")
+
+    logging.info("Loading torchscript model")
+    model = torch.jit.load(args.nn_model)
+    model.eval()
+    model.to(device)
+
+    logging.info(f"Loading HL from {args.HL}")
+    HL = kaldifst.StdVectorFst.read(args.HL)
+
+    sample_rate = 8000
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = sample_rate
+    opts.mel_opts.num_bins = 23
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {args.sound_files}")
+    waves = read_sound_files(
+        filenames=args.sound_files, expected_sample_rate=sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+
+    features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
+
+    nnet_output = model(features)
+
+    id2word = read_words(args.words)
+
+    hyps = []
+    for i in range(nnet_output.shape[0]):
+        hyp = decode(
+            filename=args.sound_files[0],
+            nnet_output=nnet_output[i],
+            HL=HL,
+            id2word=id2word,
+        )
+        hyps.append(hyp)
+
+    s = "\n"
+    for filename, hyp in zip(args.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/icefall/__init__.py b/icefall/__init__.py
index 05e2b408c..b1e4313e9 100644
--- a/icefall/__init__.py
+++ b/icefall/__init__.py
@@ -1,12 +1,6 @@
 # isort:skip_file
 
-from . import (
-    checkpoint,
-    decode,
-    dist,
-    env,
-    utils
-)
+from . import checkpoint, decode, dist, env, utils
 
 from .byte_utils import (
     byte_decode,
diff --git a/icefall/context_graph.py b/icefall/context_graph.py
index 01836df04..0b7c42c0b 100644
--- a/icefall/context_graph.py
+++ b/icefall/context_graph.py
@@ -227,7 +227,6 @@ class ContextGraph:
         filename: Optional[str] = "",
         symbol_table: Optional[Dict[int, str]] = None,
     ) -> "Digraph":  # noqa
-
         """Visualize a ContextGraph via graphviz.
 
         Render ContextGraph as an image via graphviz, and return the Digraph object;
diff --git a/icefall/ctc/.gitignore b/icefall/ctc/.gitignore
new file mode 100644
index 000000000..8154cb57f
--- /dev/null
+++ b/icefall/ctc/.gitignore
@@ -0,0 +1,2 @@
+*.pdf
+*.gv
diff --git a/icefall/ctc/README.md b/icefall/ctc/README.md
new file mode 100644
index 000000000..0096bc096
--- /dev/null
+++ b/icefall/ctc/README.md
@@ -0,0 +1,17 @@
+# Introduction
+
+This folder uses [kaldifst][kaldifst] for graph construction
+and decoders from [kaldi-decoder][kaldi-decoder] for CTC decoding.
+
+It supports only `CPU`.
+
+You can use
+
+```bash
+pip install kaldifst kaldi-decoder
+```
+to install the dependencies.
+
+[kaldi-decoder]: https://github.com/i2-fsa/kaldi-decoder
+[kaldifst]: https://github.com/k2-fsa/kaldifst
+[k2]: https://github.com/k2-fsa/k2
diff --git a/icefall/ctc/__init__.py b/icefall/ctc/__init__.py
new file mode 100644
index 000000000..b546b31af
--- /dev/null
+++ b/icefall/ctc/__init__.py
@@ -0,0 +1,6 @@
+from .prepare_lang import (
+    Lexicon,
+    make_lexicon_fst_no_silence,
+    make_lexicon_fst_with_silence,
+)
+from .topo import add_disambig_self_loops, add_one, build_standard_ctc_topo
diff --git a/icefall/ctc/prepare_lang.py b/icefall/ctc/prepare_lang.py
new file mode 100644
index 000000000..4801b1beb
--- /dev/null
+++ b/icefall/ctc/prepare_lang.py
@@ -0,0 +1,334 @@
+# Copyright    2023  Xiaomi Corp.        (author: Fangjun Kuang)
+
+"""
+The lang_dir should contain the following files:
+ - "lexicon_disambig.txt"
+ - "tokens.txt"
+ - "words.txt"
+"""
+
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import List, Tuple
+
+import kaldifst
+import re
+
+
+class Lexicon:
+    """Once constructed it is immutable"""
+
+    def __init__(
+        self,
+        lang_dir: str,
+        disambig_pattern: str = re.compile(r"^#\d+$"),
+    ):
+        """
+        Args:
+          lang_dir:
+            The path to the lang directory. We expect that it contains the
+            following files:
+              - lexicon_disambig.txt
+              - tokens.txt
+              - words.txt
+
+            The format of the above files is described below.
+
+            (1) lexicon_disambig.txt
+
+            Each line in the lexicon_disambig.txt has the following format:
+
+                word token1 token2 ... tokenN
+
+            That is, the first field is the word, the remaining fields are
+            pronunciations of this word. Fields are separated by space(s).
+
+            (2) tokens.txt
+
+            Each line in tokens.txt has two fields separated by space(s):
+
+                token ID
+
+            The first field is the token symbol and the second filed is the
+            integer ID of the token.
+
+            (3) words.txt
+
+            Each line in words.txt has two fields separated by space(s):
+
+                word ID
+
+            The first field is the word symbol and the second filed is the
+            integer ID of the word.
+          disambig_pattern:
+            It contains the pattern for disambiguation symbols.
+        """
+        lang_dir = Path(lang_dir)
+
+        lexicon_txt = lang_dir / "lexicon_disambig.txt"
+        tokens_txt = lang_dir / "tokens.txt"
+        words_txt = lang_dir / "words.txt"
+
+        assert lexicon_txt.is_file(), lexicon_txt
+        assert tokens_txt.is_file(), tokens_txt
+        assert words_txt.is_file(), words_txt
+
+        self._read_lexicon(lexicon_txt)
+        self._read_tokens(tokens_txt)
+        self._read_words(words_txt)
+
+        self.disambig_pattern = disambig_pattern
+
+        max_disambig_id = -1
+        for s, i in self.token2id.items():
+            if self.disambig_pattern.match(s) and i > max_disambig_id:
+                max_disambig_id = i
+
+        self.max_disambig_id = max_disambig_id
+
+    def _read_lexicon(self, lexicon_txt: str):
+        word2phones = defaultdict(list)
+        with open(lexicon_txt, encoding="utf-8") as f:
+            for line in f:
+                word_phones = line.strip().split()
+                assert len(word_phones) >= 2, (word_phones, line)
+                word = word_phones[0]
+                phones: str = " ".join(word_phones[1:])
+                word2phones[word].append(phones)
+                # We use a list here since a word may have multiple
+                # pronunciations
+
+        self.word2phones = word2phones
+
+    def _read_tokens(self, tokens_txt):
+        token2id = dict()
+        id2token = dict()
+        with open(tokens_txt, encoding="utf-8") as f:
+            for line in f:
+                token_id = line.strip().split()
+                assert len(token_id) == 2, token_id
+
+                token = token_id[0]
+                idx = int(token_id[1])
+
+                assert token not in token2id, f"Duplicate token {line}"
+                assert idx not in id2token, f"Duplicate ID {line}"
+
+                token2id[token] = idx
+                id2token[idx] = token
+        self.token2id = token2id
+        self.id2token = id2token
+
+    def _read_words(self, words_txt):
+        word2id = dict()
+        id2word = dict()
+        with open(words_txt, encoding="utf-8") as f:
+            for line in f:
+                word_id = line.strip().split()
+                assert len(word_id) == 2, word_id
+
+                word = word_id[0]
+                idx = int(word_id[1])
+
+                assert word not in word2id, f"Duplicate token {line}"
+                assert idx not in id2word, f"Duplicate ID {line}"
+
+                word2id[word] = idx
+                id2word[idx] = word
+
+        self.word2id = word2id
+        self.id2word = id2word
+
+    def __iter__(self) -> Tuple[str, List[str]]:
+        for word, phones_list in self.word2phones.items():
+            for phones in phones_list:
+                yield word, phones
+
+    def __str__(self):
+        return str(self.word2phones)
+
+    @property
+    def tokens(self) -> List[int]:
+        """Return a list of token IDs excluding those from
+        disambiguation symbols.
+
+        Caution:
+          0 is not a token ID so it is excluded from the return value.
+        """
+        ans = []
+        for s in self.token2id:
+            if not self.disambig_pattern.match(s):
+                ans.append(self.token2id[s])
+        if 0 in ans:
+            ans.remove(0)
+        ans.sort()
+        return ans
+
+
+# See also
+# http://vpanayotov.blogspot.com/2012/06/kaldi-decoding-graph-construction.html
+def make_lexicon_fst_with_silence(
+    lexicon: Lexicon,
+    sil_prob: float = 0.5,
+    sil_phone: str = "SIL",
+    attach_symbol_table: bool = True,
+) -> kaldifst.StdVectorFst:
+    phone2id = lexicon.token2id
+    word2id = lexicon.word2id
+
+    assert sil_phone in phone2id
+
+    assert sil_phone in phone2id, sil_phone
+
+    sil_cost = -1 * math.log(sil_prob)
+    no_sil_cost = -1 * math.log(1.0 - sil_prob)
+
+    fst = kaldifst.StdVectorFst()
+
+    start_state = fst.add_state()
+    loop_state = fst.add_state()
+    sil_state = fst.add_state()
+
+    fst.start = start_state
+    fst.set_final(state=loop_state, weight=0)
+
+    fst.add_arc(
+        state=start_state,
+        arc=kaldifst.StdArc(
+            ilabel=0,
+            olabel=0,
+            weight=no_sil_cost,
+            nextstate=loop_state,
+        ),
+    )
+
+    fst.add_arc(
+        state=start_state,
+        arc=kaldifst.StdArc(
+            ilabel=0,
+            olabel=0,
+            weight=sil_cost,
+            nextstate=sil_state,
+        ),
+    )
+
+    fst.add_arc(
+        state=sil_state,
+        arc=kaldifst.StdArc(
+            ilabel=phone2id[sil_phone],
+            olabel=0,
+            weight=0,
+            nextstate=loop_state,
+        ),
+    )
+
+    for word, phones in lexicon:
+        phoneseq = phones.split()
+        pron_cost = 0
+        cur_state = loop_state
+
+        for i in range(len(phoneseq) - 1):
+            next_state = fst.add_state()
+            fst.add_arc(
+                state=cur_state,
+                arc=kaldifst.StdArc(
+                    ilabel=phone2id[phoneseq[i]],
+                    olabel=word2id[word] if i == 0 else 0,
+                    weight=pron_cost if i == 0 else 0,
+                    nextstate=next_state,
+                ),
+            )
+            cur_state = next_state
+
+        i = len(phoneseq) - 1  # note: i == -1 if phoneseq is empty.
+
+        fst.add_arc(
+            state=cur_state,
+            arc=kaldifst.StdArc(
+                ilabel=phone2id[phoneseq[i]] if i >= 0 else 0,
+                olabel=word2id[word] if i <= 0 else 0,
+                weight=no_sil_cost + (pron_cost if i <= 0 else 0),
+                nextstate=loop_state,
+            ),
+        )
+
+        fst.add_arc(
+            state=cur_state,
+            arc=kaldifst.StdArc(
+                ilabel=phone2id[phoneseq[i]] if i >= 0 else 0,
+                olabel=word2id[word] if i <= 0 else 0,
+                weight=sil_cost + (pron_cost if i <= 0 else 0),
+                nextstate=sil_state,
+            ),
+        )
+
+    if attach_symbol_table:
+        isym = kaldifst.SymbolTable()
+        for p, i in phone2id.items():
+            isym.add_symbol(symbol=p, key=i)
+        fst.input_symbols = isym
+
+        osym = kaldifst.SymbolTable()
+        for w, i in word2id.items():
+            osym.add_symbol(symbol=w, key=i)
+        fst.output_symbols = osym
+
+    return fst
+
+
+def make_lexicon_fst_no_silence(
+    lexicon: Lexicon,
+    attach_symbol_table: bool = True,
+) -> kaldifst.StdVectorFst:
+    phone2id = lexicon.token2id
+    word2id = lexicon.word2id
+
+    fst = kaldifst.StdVectorFst()
+
+    start_state = fst.add_state()
+    fst.start = start_state
+    fst.set_final(state=start_state, weight=0)
+
+    for word, phones in lexicon:
+        phoneseq = phones.split()
+        pron_cost = 0
+        cur_state = start_state
+
+        for i in range(len(phoneseq) - 1):
+            next_state = fst.add_state()
+            fst.add_arc(
+                state=cur_state,
+                arc=kaldifst.StdArc(
+                    ilabel=phone2id[phoneseq[i]],
+                    olabel=word2id[word] if i == 0 else 0,
+                    weight=pron_cost if i == 0 else 0,
+                    nextstate=next_state,
+                ),
+            )
+            cur_state = next_state
+
+        i = len(phoneseq) - 1  # note: i == -1 if phoneseq is empty.
+
+        fst.add_arc(
+            state=cur_state,
+            arc=kaldifst.StdArc(
+                ilabel=phone2id[phoneseq[i]] if i >= 0 else 0,
+                olabel=word2id[word] if i <= 0 else 0,
+                weight=pron_cost if i <= 0 else 0,
+                nextstate=start_state,
+            ),
+        )
+
+    if attach_symbol_table:
+        isym = kaldifst.SymbolTable()
+        for p, i in phone2id.items():
+            isym.add_symbol(symbol=p, key=i)
+        fst.input_symbols = isym
+
+        osym = kaldifst.SymbolTable()
+        for w, i in word2id.items():
+            osym.add_symbol(symbol=w, key=i)
+        fst.output_symbols = osym
+
+    return fst
diff --git a/icefall/ctc/test_ctc_topo.py b/icefall/ctc/test_ctc_topo.py
new file mode 100755
index 000000000..4d4667209
--- /dev/null
+++ b/icefall/ctc/test_ctc_topo.py
@@ -0,0 +1,140 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+from pathlib import Path
+
+import graphviz
+import kaldifst
+import sentencepiece as spm
+from prepare_lang import (
+    Lexicon,
+    make_lexicon_fst_no_silence,
+    make_lexicon_fst_with_silence,
+)
+from topo import add_disambig_self_loops, add_one, build_standard_ctc_topo
+
+
+def test_yesno():
+    lang_dir = "/Users/fangjun/open-source/icefall/egs/yesno/ASR/data/lang_phone"
+    if not Path(lang_dir).is_dir():
+        print(f"{lang_dir} does not exist! Skip testing")
+        return
+    lexicon = Lexicon(lang_dir)
+    max_token_id = max(lexicon.tokens)
+
+    H = build_standard_ctc_topo(max_token_id=max_token_id)
+
+    isym = kaldifst.SymbolTable()
+    isym.add_symbol(symbol="<blk>", key=0)
+    for i in range(1, max_token_id + 1):
+        isym.add_symbol(symbol=lexicon.id2token[i], key=i)
+
+    osym = kaldifst.SymbolTable()
+    osym.add_symbol(symbol="<eps>", key=0)
+    for i in range(1, max_token_id + 1):
+        osym.add_symbol(symbol=lexicon.id2token[i], key=i)
+
+    H.input_symbols = isym
+    H.output_symbols = osym
+
+    fst_dot = kaldifst.draw(H, acceptor=False, portrait=True)
+    source = graphviz.Source(fst_dot)
+    source.render(outfile="standard_ctc_topo_yesno.pdf")
+    # See the link below to visualize the above PDF
+    # https://t.ly/7uXZ9
+
+    # Now test HL
+
+    # We need to add one to all tokens since we want to use ID 0
+    # for epsilon
+    add_one(H, treat_ilabel_zero_specially=False, update_olabel=True)
+
+    add_disambig_self_loops(
+        H,
+        start=lexicon.token2id["#0"] + 1,
+        end=lexicon.max_disambig_id,
+    )
+
+    fst_dot = kaldifst.draw(H, acceptor=False, portrait=True)
+    source = graphviz.Source(fst_dot)
+    source.render(outfile="standard_ctc_topo_disambig_yesno.pdf")
+
+    L = make_lexicon_fst_with_silence(lexicon)
+
+    # We also need to change the input labels of L
+    add_one(L, treat_ilabel_zero_specially=True, update_olabel=False)
+
+    H.output_symbols = None
+
+    kaldifst.arcsort(H, sort_type="olabel")
+    kaldifst.arcsort(L, sort_type="ilabel")
+    HL = kaldifst.compose(H, L)
+
+    lexicon.id2token[0] = "<blk>"
+    lexicon.token2id["<blk>"] = 0
+
+    isym = kaldifst.SymbolTable()
+    isym.add_symbol(symbol="<eps>", key=0)
+    for i in range(0, lexicon.max_disambig_id + 1):
+        isym.add_symbol(symbol=lexicon.id2token[i], key=i + 1)
+
+    osym = kaldifst.SymbolTable()
+    for i, word in lexicon.id2word.items():
+        osym.add_symbol(symbol=word, key=i)
+
+    HL.input_symbols = isym
+    HL.output_symbols = osym
+
+    fst_dot = kaldifst.draw(HL, acceptor=False, portrait=True)
+    source = graphviz.Source(fst_dot)
+    source.render(outfile="HL_yesno.pdf")
+
+
+def test_librispeech():
+    lang_dir = (
+        "/star-fj/fangjun/open-source/icefall-2/egs/librispeech/ASR/data/lang_bpe_500"
+    )
+
+    if not Path(lang_dir).is_dir():
+        print(f"{lang_dir} does not exist! Skip testing")
+        return
+
+    lexicon = Lexicon(lang_dir)
+    HL = kaldifst.StdVectorFst.read(lang_dir + "/HL.fst")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(lang_dir + "/bpe.model")
+
+    i = lexicon.word2id["HELLOA"]
+    k = lexicon.word2id["WORLD"]
+    print(i, k)
+    s = f"""
+        0 1 {i} {i}
+        1 2 {k} {k}
+        2
+    """
+    fst = kaldifst.compile(
+        s=s,
+        acceptor=False,
+    )
+
+    L = make_lexicon_fst_no_silence(lexicon, attach_symbol_table=False)
+    kaldifst.arcsort(L, sort_type="olabel")
+    with open("L.fst.txt", "w") as f:
+        print(L, file=f)
+
+    fst = kaldifst.compose(L, fst)
+    print(fst)
+    fst_dot = kaldifst.draw(fst, acceptor=False, portrait=True)
+    source = graphviz.Source(fst_dot)
+    source.render(outfile="a.pdf")
+    print(sp.encode(["HELLOA", "WORLD"]))
+
+
+def main():
+    test_yesno()
+    test_librispeech()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/ctc/test_prepare_lang.py b/icefall/ctc/test_prepare_lang.py
new file mode 100755
index 000000000..6c4b9e510
--- /dev/null
+++ b/icefall/ctc/test_prepare_lang.py
@@ -0,0 +1,43 @@
+#!/usr/bin/env python3
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+from pathlib import Path
+
+import graphviz
+import kaldifst
+from prepare_lang import Lexicon, make_lexicon_fst_with_silence
+
+
+def test_yesno():
+    lang_dir = "/Users/fangjun/open-source/icefall/egs/yesno/ASR/data/lang_phone"
+    if not Path(lang_dir).is_dir():
+        print(f"{lang_dir} does not exist! Skip testing")
+        return
+
+    lexicon = Lexicon(lang_dir)
+
+    L = make_lexicon_fst_with_silence(lexicon)
+
+    isym = kaldifst.SymbolTable()
+    for i, token in lexicon.id2token.items():
+        isym.add_symbol(symbol=token, key=i)
+
+    osym = kaldifst.SymbolTable()
+    for i, word in lexicon.id2word.items():
+        osym.add_symbol(symbol=word, key=i)
+
+    L.input_symbols = isym
+    L.output_symbols = osym
+    fst_dot = kaldifst.draw(L, acceptor=False, portrait=True)
+    source = graphviz.Source(fst_dot)
+    source.render(outfile="L_yesno.pdf")
+    # See the link below to visualize the above PDF
+    # https://t.ly/jMfXW
+
+
+def main():
+    test_yesno()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/ctc/topo.py b/icefall/ctc/topo.py
new file mode 100644
index 000000000..6a96dd038
--- /dev/null
+++ b/icefall/ctc/topo.py
@@ -0,0 +1,137 @@
+# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+import kaldifst
+
+
+# Note the name contains `standard`; it means there will be non-standard
+# topologies.
+def build_standard_ctc_topo(max_token_id: int) -> kaldifst.StdVectorFst:
+    """Build a standard CTC topology.
+
+    Args:
+      Maximum valid token ID. We assume token IDs are contiguous
+      and starts from 0. In other words, the vocabulary size is
+      ``max_token_id + 1``. We assume the ID of the blank symbol is 0.
+    """
+    # Token ID starts from 0 and there are as many states as the
+    # number of tokens.
+    #
+    # Note that epsilon is not a token and the token with ID 0 in tokens.txt
+    # is not an epsilon. It means input label 0 of the resulting FST does
+    # not represent an epsilon.
+    #
+    # You can use the function `add_one()` to modify the input/output labels
+    # of the resulting FST
+
+    num_states = max_token_id + 1
+
+    # Step 1: Create as many states as the number of tokens.
+    # Each state is a final state
+    fst = kaldifst.StdVectorFst()
+    for i in range(num_states):
+        s = fst.add_state()
+        fst.set_final(state=s, weight=0)
+
+    # Step 2: Set state 0 as the start state.
+    # We assume the ID of the blank symbol is 0.
+    fst.start = 0
+
+    # Step 3: Build a fully connected graph.
+    for i in range(num_states):
+        for k in range(num_states):
+            fst.add_arc(
+                state=i,
+                arc=kaldifst.StdArc(
+                    ilabel=k,
+                    olabel=k if i != k else 0,  # if i==k, it is a self loop
+                    weight=0,
+                    nextstate=k,
+                ),
+            )
+    # Please see ./test_ctc_topo.py if you want to know what the resulting
+    # FST looks like
+
+    return fst
+
+
+def add_one(
+    fst: kaldifst.StdVectorFst,
+    treat_ilabel_zero_specially: bool,
+    update_olabel: bool,
+) -> None:
+    """Modify the input and output labels of the given FST in-place.
+
+    Args:
+      fst:
+        The FST to be modified. It is changed in-place.
+      treat_ilabel_zero_specially:
+        If True, then every non-zero input label is increased by one and the
+        zero input label is not changed.
+        If False, then every input label is increased by one.
+      update_olabel:
+        If False, the output label is not changed.
+        If True, then every non-zero output label is increased by one.
+        In either case, output label with 0 is not changed.
+    """
+    for state in kaldifst.StateIterator(fst):
+        for arc in kaldifst.ArcIterator(fst, state):
+            # If treat_ilabel_zero_specially is False, we always change it
+            # Otherwise, we only change non-zero input labels
+            if treat_ilabel_zero_specially is False or arc.ilabel != 0:
+                arc.ilabel += 1
+
+            if update_olabel and arc.olabel != 0:
+                arc.olabel += 1
+
+    if fst.input_symbols is not None:
+        input_symbols = kaldifst.SymbolTable()
+        input_symbols.add_symbol(symbol="<eps>", key=0)
+
+        for i in range(0, fst.input_symbols.num_symbols()):
+            s = fst.input_symbols.find(i)
+            input_symbols.add_symbol(symbol=s, key=i + 1)
+
+        fst.input_symbols = input_symbols
+
+    if update_olabel and fst.output_symbols is not None:
+        output_symbols = kaldifst.SymbolTable()
+        output_symbols.add_symbol(symbol="<eps>", key=0)
+
+        for i in range(0, fst.output_symbols.num_symbols()):
+            s = fst.output_symbols.find(i)
+            output_symbols.add_symbol(symbol=s, key=i + 1)
+
+        fst.output_symbols = output_symbols
+
+
+def add_disambig_self_loops(fst: kaldifst.StdVectorFst, start: int, end: int):
+    """Add self-loops to each state.
+
+    For each disambig symbol, we add a self-loop with input label disambig_id
+    and output label diambig_id of that disambig symbol.
+
+    Args:
+      fst:
+        It is changed in-place.
+      start:
+        The ID of #0
+      end:
+        The ID of the last disambig symbol. For instance if there are 3
+        disambig symbols ``#0``, ``#1``, and ``#2``, then ``end`` is the ID
+        of ``#2``.
+    """
+    for state in kaldifst.StateIterator(fst):
+        for i in range(start, end + 1):
+            fst.add_arc(
+                state=state,
+                arc=kaldifst.StdArc(
+                    ilabel=i,
+                    olabel=i,
+                    weight=0,
+                    nextstate=state,
+                ),
+            )
+
+    if fst.output_symbols:
+        for i in range(start, end + 1):
+            fst.output_symbols.add_symbol(symbol=f"#{i-start}", key=i)
diff --git a/icefall/diagnostics.py b/icefall/diagnostics.py
index 98870684e..ebf61784e 100644
--- a/icefall/diagnostics.py
+++ b/icefall/diagnostics.py
@@ -23,6 +23,7 @@ from typing import Optional, Tuple, List
 import torch
 from torch import Tensor, nn
 
+
 class TensorDiagnosticOptions(object):
     """Options object for tensor diagnostics:
 
@@ -77,11 +78,11 @@ def get_tensor_stats(
     elif stats_type == "abs":
         x = x.abs()
     elif stats_type == "rms":
-        x = x ** 2
+        x = x**2
     elif stats_type == "positive":
         x = (x > 0).to(dtype=torch.float)
     else:
-        assert stats_type in [ "value", "max", "min" ]
+        assert stats_type in ["value", "max", "min"]
 
     sum_dims = [d for d in range(x.ndim) if d != dim]
     if len(sum_dims) > 0:
@@ -121,10 +122,10 @@ class TensorDiagnostic(object):
         self.class_name = None  # will assign in accumulate()
 
         self.stats = None  # we'll later assign a list to self.stats.
-                           # It's a list of dicts, indexed by dim (i.e. by the
-                           # axis of the tensor).  The dicts, in turn, are
-                           # indexed by `stats-type` which are strings in
-                           # ["abs", "max", "min", "positive", "value", "rms"].
+        # It's a list of dicts, indexed by dim (i.e. by the
+        # axis of the tensor).  The dicts, in turn, are
+        # indexed by `stats-type` which are strings in
+        # ["abs", "max", "min", "positive", "value", "rms"].
 
         # scalar_stats contains some analysis of the activations and gradients,
         self.scalar_stats = None
@@ -139,7 +140,6 @@ class TensorDiagnostic(object):
         # only adding a new element to the list if there was a different dim.
         # if the string in the key is "eigs", if we detect a length mismatch we put None as the value.
 
-
     def accumulate(self, x, class_name: Optional[str] = None):
         """
         Accumulate tensors.
@@ -193,17 +193,12 @@ class TensorDiagnostic(object):
                         done = True
                         break
                 if not done:
-                    if (
-                        this_dim_stats[stats_type] != []
-                        and stats_type == "eigs"
-                    ):
+                    if this_dim_stats[stats_type] != [] and stats_type == "eigs":
                         # >1 size encountered on this dim, e.g. it's a batch or time dimension,
                         # don't accumulat "eigs" stats type, it uses too much memory
                         this_dim_stats[stats_type] = None
                     else:
-                        this_dim_stats[stats_type].append(
-                            TensorAndCount(stats, count)
-                        )
+                        this_dim_stats[stats_type].append(TensorAndCount(stats, count))
 
     def print_diagnostics(self):
         """Print diagnostics for each dimension of the tensor."""
@@ -220,8 +215,11 @@ class TensorDiagnostic(object):
                     for r, v in zip(rms_stats_list, value_stats_list):
                         stddev_stats_list.append(
                             # r.count and v.count should be the same, but we don't check this.
-                            TensorAndCount(r.tensor - v.tensor * v.tensor / (v.count + 1.0e-20),
-                                           r.count))
+                            TensorAndCount(
+                                r.tensor - v.tensor * v.tensor / (v.count + 1.0e-20),
+                                r.count,
+                            )
+                        )
                     this_dim_stats["stddev"] = stddev_stats_list
 
             for stats_type, stats_list in this_dim_stats.items():
@@ -232,7 +230,6 @@ class TensorDiagnostic(object):
                     assert stats_type == "eigs"
                     continue
 
-
                 def get_count(count):
                     return 1 if stats_type in ["max", "min"] else count
 
@@ -247,25 +244,33 @@ class TensorDiagnostic(object):
 
                 if stats_type == "eigs":
                     try:
-                        eigs, _ = torch.symeig(stats)
+                        if hasattr(torch, 'linalg') and hasattr(torch.linalg, 'eigh'):
+                            eigs, _ = torch.linalg.eigh(stats)
+                        else:
+                            eigs, _ = torch.symeig(stats)
                         stats = eigs.abs().sqrt()
                     except:  # noqa
                         print(
                             "Error getting eigenvalues, trying another method."
                         )
-                        eigs, _ = torch.eig(stats)
-                        stats = eigs.norm(dim=1).sqrt()
+                        if hasattr(torch, 'linalg') and hasattr(torch.linalg, 'eig'):
+                            eigs, _ = torch.linalg.eig(stats)
+                            eigs = eigs.abs()
+                        else:
+                            eigs, _ = torch.eig(stats)
+                            eigs = eigs.norm(dim=1)
+                        stats = eigs.sqrt()
                         # sqrt so it reflects data magnitude, like stddev- not variance
 
-                if stats_type in [ "rms", "stddev" ]:
+                if stats_type in ["rms", "stddev"]:
                     # we stored the square; after aggregation we need to take sqrt.
                     stats = stats.sqrt()
 
                 # if `summarize` we print percentiles of the stats; else,
                 # we print out individual elements.
-                summarize = (
-                    len(stats_list) > 1
-                ) or self.opts.dim_is_summarized(stats.numel())
+                summarize = (len(stats_list) > 1) or self.opts.dim_is_summarized(
+                    stats.numel()
+                )
                 if summarize:  # usually `summarize` will be true
                     # print out percentiles.
                     stats = stats.sort()[0]
@@ -282,15 +287,15 @@ class TensorDiagnostic(object):
                     ans = stats.tolist()
                     ans = ["%.2g" % x for x in ans]
                     ans = "[" + " ".join(ans) + "]"
-                if stats_type in [ "value", "rms", "stddev", "eigs" ]:
+                if stats_type in ["value", "rms", "stddev", "eigs"]:
                     # This norm is useful because it is strictly less than the largest
                     # sqrt(eigenvalue) of the variance, which we print out, and shows,
                     # speaking in an approximate way, how much of that largest eigenvalue
                     # can be attributed to the mean of the distribution.
-                    norm = (stats ** 2).sum().sqrt().item()
+                    norm = (stats**2).sum().sqrt().item()
                     ans += f", norm={norm:.2g}"
                 mean = stats.mean().item()
-                rms = (stats ** 2).mean().sqrt().item()
+                rms = (stats**2).mean().sqrt().item()
                 ans += f", mean={mean:.3g}, rms={rms:.3g}"
 
                 # OK, "ans" contains the actual stats, e.g.
@@ -298,11 +303,11 @@ class TensorDiagnostic(object):
 
                 sizes = [x.tensor.shape[0] for x in stats_list]
                 size_str = (
-                    f"{sizes[0]}"
-                    if len(sizes) == 1
-                    else f"{min(sizes)}..{max(sizes)}"
+                    f"{sizes[0]}" if len(sizes) == 1 else f"{min(sizes)}..{max(sizes)}"
+                )
+                maybe_class_name = (
+                    f" type={self.class_name}," if self.class_name is not None else ""
                 )
-                maybe_class_name = f" type={self.class_name}," if self.class_name is not None else ""
                 print(
                     f"module={self.name},{maybe_class_name} dim={dim}, size={size_str}, {stats_type} {ans}"
                 )
@@ -330,7 +335,6 @@ class ScalarDiagnostic(object):
         self.sum_gradsq = None
         self.sum_abs_grad = None
 
-
     def accumulate_input(self, x: Tensor, class_name: Optional[str] = None):
         """
         Called in forward pass.
@@ -347,8 +351,10 @@ class ScalarDiagnostic(object):
 
         limit = 10
         if len(self.saved_inputs) > limit:
-            print(f"ERROR: forward pass called for this module over {limit} times with no backward pass. "
-                  f" Will not accumulate scalar stats.")
+            print(
+                f"ERROR: forward pass called for this module over {limit} times with no backward pass. "
+                f" Will not accumulate scalar stats."
+            )
             self.is_ok = False
             return
         self.saved_inputs.append(x)
@@ -359,11 +365,15 @@ class ScalarDiagnostic(object):
         if self.is_forward_pass:
             self.is_forward_pass = False
 
-        last_shape = 'n/a' if len(self.saved_inputs) == 0 else self.saved_inputs[-1].shape
+        last_shape = (
+            "n/a" if len(self.saved_inputs) == 0 else self.saved_inputs[-1].shape
+        )
         if len(self.saved_inputs) == 0 or grad.shape != last_shape:
-            print(f"ERROR: shape mismatch or no forward activation present when backward "
-                  f"pass called: grad shape ={tuple(grad.shape)}, num-saved-inputs={len(self.saved_inputs)}"
-                  f", shape-of-last-saved-input={last_shape}")
+            print(
+                f"ERROR: shape mismatch or no forward activation present when backward "
+                f"pass called: grad shape ={tuple(grad.shape)}, num-saved-inputs={len(self.saved_inputs)}"
+                f", shape-of-last-saved-input={last_shape}"
+            )
             self.is_ok = False
             return
 
@@ -384,11 +394,19 @@ class ScalarDiagnostic(object):
             self.tick_scale = float(x_abs_sorted[index] / num_ticks_per_side)
 
             # integerize from tick * (-num ticks_per_side ..  num_ticks_per_side - 1]
-            self.counts = torch.zeros(2 * num_ticks_per_side, dtype=torch.long, device=x.device)
-            self.sum_grad = torch.zeros(2 * num_ticks_per_side, dtype=torch.double, device=x.device)
+            self.counts = torch.zeros(
+                2 * num_ticks_per_side, dtype=torch.long, device=x.device
+            )
+            self.sum_grad = torch.zeros(
+                2 * num_ticks_per_side, dtype=torch.double, device=x.device
+            )
             # sum_gradsq is for getting error bars.
-            self.sum_gradsq = torch.zeros(2 * num_ticks_per_side, dtype=torch.double, device=x.device)
-            self.sum_abs_grad = torch.zeros(2 * num_ticks_per_side, dtype=torch.double, device=x.device)
+            self.sum_gradsq = torch.zeros(
+                2 * num_ticks_per_side, dtype=torch.double, device=x.device
+            )
+            self.sum_abs_grad = torch.zeros(
+                2 * num_ticks_per_side, dtype=torch.double, device=x.device
+            )
 
         # this will round down.
         x = (x / self.tick_scale).to(torch.long)
@@ -397,20 +415,21 @@ class ScalarDiagnostic(object):
 
         self.counts.index_add_(dim=0, index=x, source=torch.ones_like(x))
         self.sum_grad.index_add_(dim=0, index=x, source=grad.to(torch.double))
-        self.sum_gradsq.index_add_(dim=0, index=x, source=(grad*grad).to(torch.double))
+        self.sum_gradsq.index_add_(
+            dim=0, index=x, source=(grad * grad).to(torch.double)
+        )
         self.sum_abs_grad.index_add_(dim=0, index=x, source=grad.abs().to(torch.double))
 
-
     def print_diagnostics(self):
         """Print diagnostics."""
         if self.is_ok is False or self.counts is None:
             print(f"Warning: no stats accumulated for {self.name}, is_ok={self.is_ok}")
             return
 
-        counts = self.counts.to('cpu')
-        sum_grad = self.sum_grad.to(device='cpu', dtype=torch.float32)
-        sum_gradsq = self.sum_gradsq.to(device='cpu', dtype=torch.float32)
-        sum_abs_grad = self.sum_abs_grad.to(device='cpu', dtype=torch.float32)
+        counts = self.counts.to("cpu")
+        sum_grad = self.sum_grad.to(device="cpu", dtype=torch.float32)
+        sum_gradsq = self.sum_gradsq.to(device="cpu", dtype=torch.float32)
+        sum_abs_grad = self.sum_abs_grad.to(device="cpu", dtype=torch.float32)
 
         counts_cumsum = counts.cumsum(dim=0)
         counts_tot = counts_cumsum[-1]
@@ -433,19 +452,22 @@ class ScalarDiagnostic(object):
         bin_abs_grad = torch.zeros(num_bins)
         bin_abs_grad.index_add_(dim=0, index=bin_indexes, source=sum_abs_grad)
 
-        avg_grad = (bin_grad / bin_counts)
+        avg_grad = bin_grad / bin_counts
         avg_grad_stddev = (bin_gradsq / bin_counts).sqrt()
 
-        bin_boundary_counts = torch.arange(num_bins + 1, dtype=torch.long) * counts_per_bin
+        bin_boundary_counts = (
+            torch.arange(num_bins + 1, dtype=torch.long) * counts_per_bin
+        )
         bin_tick_indexes = torch.searchsorted(counts_cumsum, bin_boundary_counts)
         # boundaries are the "x" values between the bins, e.g. corresponding to the
         # locations of percentiles of the distribution.
         num_ticks_per_side = counts.numel() // 2
         bin_boundaries = (bin_tick_indexes - num_ticks_per_side) * self.tick_scale
 
-
         bin_grad = bin_grad / (bin_counts + 1)
-        bin_conf_interval = bin_gradsq.sqrt() / (bin_counts + 1) # consider this a standard deviation.
+        bin_conf_interval = bin_gradsq.sqrt() / (
+            bin_counts + 1
+        )  # consider this a standard deviation.
         # bin_grad / bin_abs_grad will give us a sense for how important in a practical sense,
         # the gradients are.
         bin_abs_grad = bin_abs_grad / (bin_counts + 1)
@@ -458,8 +480,9 @@ class ScalarDiagnostic(object):
             x = "[" + " ".join(x) + "]"
             return x
 
-
-        maybe_class_name = f" type={self.class_name}," if self.class_name is not None else ""
+        maybe_class_name = (
+            f" type={self.class_name}," if self.class_name is not None else ""
+        )
 
         print(
             f"module={self.name},{maybe_class_name} bin-boundaries={tensor_to_str(bin_boundaries)}, "
@@ -467,7 +490,6 @@ class ScalarDiagnostic(object):
         )
 
 
-
 class ModelDiagnostic(object):
     """This class stores diagnostics for all tensors in the torch.nn.Module.
 
@@ -485,9 +507,8 @@ class ModelDiagnostic(object):
             self.opts = opts
         self.diagnostics = dict()
 
-
     def __getitem__(self, name: str):
-        T = ScalarDiagnostic if name[-7:] == '.scalar' else TensorDiagnostic
+        T = ScalarDiagnostic if name[-7:] == ".scalar" else TensorDiagnostic
         if name not in self.diagnostics:
             self.diagnostics[name] = T(self.opts, name)
         return self.diagnostics[name]
@@ -502,18 +523,19 @@ def get_class_name(module: nn.Module):
     ans = type(module).__name__
     # we put the below in try blocks in case anyone is using a different version of these modules that
     # might have different member names.
-    if ans == 'Balancer' or ans == 'ActivationBalancer':
+    if ans == "Balancer" or ans == "ActivationBalancer":
         try:
-            ans += f'[{float(module.min_positive)},{float(module.max_positive)},{float(module.min_abs)},{float(module.max_abs)}]'
+            ans += f"[{float(module.min_positive)},{float(module.max_positive)},{float(module.min_abs)},{float(module.max_abs)}]"
         except:
             pass
-    elif ans == 'AbsValuePenalizer':
+    elif ans == "AbsValuePenalizer":
         try:
-            ans += f'[{module.limit}]'
+            ans += f"[{module.limit}]"
         except:
             pass
     return ans
 
+
 def attach_diagnostics(
     model: nn.Module, opts: Optional[TensorDiagnosticOptions] = None
 ) -> ModelDiagnostic:
@@ -538,73 +560,82 @@ def attach_diagnostics(
         if name == "":
             name = "<top-level>"
 
-
-
         # Setting model_diagnostic=ans and n=name below, instead of trying to
         # capture the variables, ensures that we use the current values.
         # (this matters for `name`, since the variable gets overwritten).
         # These closures don't really capture by value, only by
         # "the final value the variable got in the function" :-(
-        def forward_hook(
-            _module, _input, _output, _model_diagnostic=ans, _name=name
-        ):
+        def forward_hook(_module, _input, _output, _model_diagnostic=ans, _name=name):
             if isinstance(_output, tuple) and len(_output) == 1:
                 _output = _output[0]
 
-            if isinstance(_output, Tensor) and _output.dtype in ( torch.float32, torch.float16, torch.float64 ):
-                _model_diagnostic[f"{_name}.output"].accumulate(_output,
-                                                                class_name=get_class_name(_module))
+            if isinstance(_output, Tensor) and _output.dtype in (
+                torch.float32,
+                torch.float16,
+                torch.float64,
+            ):
+                _model_diagnostic[f"{_name}.output"].accumulate(
+                    _output, class_name=get_class_name(_module)
+                )
             elif isinstance(_output, tuple):
                 for i, o in enumerate(_output):
-                    if o.dtype in ( torch.float32, torch.float16, torch.float64 ):
+                    if isinstance(o, Tensor) and o.dtype in ( torch.float32, torch.float16, torch.float64 ):
                         _model_diagnostic[f"{_name}.output[{i}]"].accumulate(o,
                                                                              class_name=get_class_name(_module))
-
-        def backward_hook(
-            _module, _input, _output, _model_diagnostic=ans, _name=name
-        ):
+                        
+        def backward_hook(_module, _input, _output, _model_diagnostic=ans, _name=name):
             if isinstance(_output, tuple) and len(_output) == 1:
                 _output = _output[0]
-            if isinstance(_output, Tensor) and _output.dtype in ( torch.float32, torch.float16, torch.float64 ):
-                _model_diagnostic[f"{_name}.grad"].accumulate(_output,
-                                                              class_name=get_class_name(_module))
+            if isinstance(_output, Tensor) and _output.dtype in (
+                torch.float32,
+                torch.float16,
+                torch.float64,
+            ):
+                _model_diagnostic[f"{_name}.grad"].accumulate(
+                    _output, class_name=get_class_name(_module)
+                )
             elif isinstance(_output, tuple):
                 for i, o in enumerate(_output):
-                    if o.dtype in ( torch.float32, torch.float16, torch.float64 ):
+                    if isinstance(o, Tensor) and o.dtype in ( torch.float32, torch.float16, torch.float64 ):
                         _model_diagnostic[f"{_name}.grad[{i}]"].accumulate(o,
                                                                            class_name=get_class_name(_module))
-
-
         module.register_forward_hook(forward_hook)
         module.register_backward_hook(backward_hook)
 
-        if type(module).__name__ in  ["Sigmoid", "Tanh", "ReLU", "TanSwish", "Swish", "DoubleSwish", "Swoosh"]:
+        if type(module).__name__ in [
+            "Sigmoid",
+            "Tanh",
+            "ReLU",
+            "TanSwish",
+            "Swish",
+            "DoubleSwish",
+            "Swoosh",
+        ]:
             # For these specific module types, accumulate some additional diagnostics
             # that can help us improve the activation function.  These require a lot of memory,
             # to save the forward activations, so limit this to some select classes.
             # Note: this will not work correctly for all model types.
             def scalar_forward_hook(
-                    _module, _input, _output, _model_diagnostic=ans, _name=name
+                _module, _input, _output, _model_diagnostic=ans, _name=name
             ):
                 if isinstance(_input, tuple):
-                    _input, = _input
+                    (_input,) = _input
                 assert isinstance(_input, Tensor)
-                _model_diagnostic[f"{_name}.scalar"].accumulate_input(_input,
-                                                                      class_name=get_class_name(_module))
+                _model_diagnostic[f"{_name}.scalar"].accumulate_input(
+                    _input, class_name=get_class_name(_module)
+                )
 
             def scalar_backward_hook(
-                    _module, _input, _output, _model_diagnostic=ans, _name=name
+                _module, _input, _output, _model_diagnostic=ans, _name=name
             ):
                 if isinstance(_output, tuple):
-                    _output, = _output
+                    (_output,) = _output
                 assert isinstance(_output, Tensor)
                 _model_diagnostic[f"{_name}.scalar"].accumulate_output_grad(_output)
 
             module.register_forward_hook(scalar_forward_hook)
             module.register_backward_hook(scalar_backward_hook)
 
-
-
     for name, parameter in model.named_parameters():
 
         def param_backward_hook(
diff --git a/icefall/otc_graph_compiler.py b/icefall/otc_graph_compiler.py
new file mode 100644
index 000000000..bfd679452
--- /dev/null
+++ b/icefall/otc_graph_compiler.py
@@ -0,0 +1,246 @@
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#                2023  Johns Hopkins University (author: Dongji Gao)
+#
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from pathlib import Path
+from typing import List, Union
+
+import k2
+import sentencepiece as spm
+import torch
+
+from icefall.utils import str2bool
+
+
+class OtcTrainingGraphCompiler(object):
+    def __init__(
+        self,
+        lang_dir: Path,
+        otc_token: str,
+        device: Union[str, torch.device] = "cpu",
+        sos_token: str = "<sos/eos>",
+        eos_token: str = "<sos/eos>",
+        initial_bypass_weight: float = 0.0,
+        initial_self_loop_weight: float = 0.0,
+        bypass_weight_decay: float = 0.0,
+        self_loop_weight_decay: float = 0.0,
+    ) -> None:
+        """
+        Args:
+          lang_dir:
+            This directory is expected to contain the following files:
+
+                - bpe.model
+                - words.txt
+          otc_token:
+            The special token in OTC that represent all non-blank tokens
+          device:
+            It indicates CPU or CUDA.
+          sos_token:
+            The word piece that represents sos.
+          eos_token:
+            The word piece that represents eos.
+        """
+        lang_dir = Path(lang_dir)
+        bpe_model_file = lang_dir / "bpe.model"
+        sp = spm.SentencePieceProcessor()
+        sp.load(str(bpe_model_file))
+        self.sp = sp
+        self.token_table = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+
+        self.otc_token = otc_token
+        assert self.otc_token in self.token_table
+
+        self.device = device
+
+        self.sos_id = self.sp.piece_to_id(sos_token)
+        self.eos_id = self.sp.piece_to_id(eos_token)
+
+        assert self.sos_id != self.sp.unk_id()
+        assert self.eos_id != self.sp.unk_id()
+
+        max_token_id = self.get_max_token_id()
+        ctc_topo = k2.ctc_topo(max_token_id, modified=False)
+        self.ctc_topo = ctc_topo.to(self.device)
+
+        self.initial_bypass_weight = initial_bypass_weight
+        self.initial_self_loop_weight = initial_self_loop_weight
+        self.bypass_weight_decay = bypass_weight_decay
+        self.self_loop_weight_decay = self_loop_weight_decay
+
+    def get_max_token_id(self):
+        max_token_id = 0
+        for symbol in self.token_table.symbols:
+            if not symbol.startswith("#"):
+                max_token_id = max(self.token_table[symbol], max_token_id)
+        assert max_token_id > 0
+
+        return max_token_id
+
+    def make_arc(
+        self,
+        from_state: int,
+        to_state: int,
+        symbol: Union[str, int],
+        weight: float,
+    ):
+        return f"{from_state} {to_state} {symbol} {weight}"
+
+    def texts_to_ids(self, texts: List[str]) -> List[List[int]]:
+        """Convert a list of texts to a list-of-list of piece IDs.
+
+        Args:
+          texts:
+            It is a list of strings. Each string consists of space(s)
+            separated words. An example containing two strings is given below:
+
+                ['HELLO ICEFALL', 'HELLO k2']
+        Returns:
+          Return a list-of-list of piece IDs.
+        """
+        return self.sp.encode(texts, out_type=int)
+
+    def compile(
+        self,
+        texts: List[str],
+        allow_bypass_arc: str2bool = True,
+        allow_self_loop_arc: str2bool = True,
+        bypass_weight: float = 0.0,
+        self_loop_weight: float = 0.0,
+    ) -> k2.Fsa:
+        """Build a OTC graph from a texts (list of words).
+
+        Args:
+          texts:
+            A list of strings. Each string contains a sentence for an utterance.
+            A sentence consists of spaces separated words. An example `texts`
+            looks like:
+              ['hello icefall', 'CTC training with k2']
+          allow_bypass_arc:
+            Whether to add bypass arc to training graph for substitution
+            and insertion errors (wrong or extra words in the transcript).
+          allow_self_loop_arc:
+            Whether to add self-loop arc to training graph for deletion
+            errors (missing words in the transcript).
+          bypass_weight:
+            Weight associated with bypass arc.
+          self_loop_weight:
+            Weight associated with self-loop arc.
+
+        Return:
+          Return an FsaVec, which is the result of composing a
+          CTC topology with OTC FSAs constructed from the given texts.
+        """
+
+        transcript_fsa = self.convert_transcript_to_fsa(
+            texts,
+            self.otc_token,
+            allow_bypass_arc,
+            allow_self_loop_arc,
+            bypass_weight,
+            self_loop_weight,
+        )
+        transcript_fsa = transcript_fsa.to(self.device)
+        fsa_with_self_loop = k2.remove_epsilon_and_add_self_loops(transcript_fsa)
+        fsa_with_self_loop = k2.arc_sort(fsa_with_self_loop)
+
+        graph = k2.compose(
+            self.ctc_topo,
+            fsa_with_self_loop,
+            treat_epsilons_specially=False,
+        )
+        assert graph.requires_grad is False
+
+        return graph
+
+    def convert_transcript_to_fsa(
+        self,
+        texts: List[str],
+        otc_token: str,
+        allow_bypass_arc: str2bool = True,
+        allow_self_loop_arc: str2bool = True,
+        bypass_weight: float = 0.0,
+        self_loop_weight: float = 0.0,
+    ):
+        otc_token_id = self.token_table[otc_token]
+
+        transcript_fsa_list = []
+        for text in texts:
+            text_piece_ids = []
+
+            for word in text.split():
+                piece_ids = self.sp.encode(word, out_type=int)
+                text_piece_ids.append(piece_ids)
+
+            arcs = []
+            start_state = 0
+            cur_state = start_state
+            next_state = 1
+
+            for piece_ids in text_piece_ids:
+                bypass_cur_state = cur_state
+
+                if allow_self_loop_arc:
+                    self_loop_arc = self.make_arc(
+                        cur_state,
+                        cur_state,
+                        otc_token_id,
+                        self_loop_weight,
+                    )
+                    arcs.append(self_loop_arc)
+
+                for piece_id in piece_ids:
+                    arc = self.make_arc(cur_state, next_state, piece_id, 0.0)
+                    arcs.append(arc)
+
+                    cur_state = next_state
+                    next_state += 1
+
+                bypass_next_state = cur_state
+                if allow_bypass_arc:
+                    bypass_arc = self.make_arc(
+                        bypass_cur_state,
+                        bypass_next_state,
+                        otc_token_id,
+                        bypass_weight,
+                    )
+                    arcs.append(bypass_arc)
+                bypass_cur_state = cur_state
+
+            if allow_self_loop_arc:
+                self_loop_arc = self.make_arc(
+                    cur_state,
+                    cur_state,
+                    otc_token_id,
+                    self_loop_weight,
+                )
+                arcs.append(self_loop_arc)
+
+            # Deal with final state
+            final_state = next_state
+            final_arc = self.make_arc(cur_state, final_state, -1, 0.0)
+            arcs.append(final_arc)
+            arcs.append(f"{final_state}")
+            sorted_arcs = sorted(arcs, key=lambda a: int(a.split()[0]))
+
+            transcript_fsa = k2.Fsa.from_str("\n".join(sorted_arcs))
+            transcript_fsa = k2.arc_sort(transcript_fsa)
+            transcript_fsa_list.append(transcript_fsa)
+
+        transcript_fsa_vec = k2.create_fsa_vec(transcript_fsa_list)
+
+        return transcript_fsa_vec
diff --git a/icefall/profiler.py b/icefall/profiler.py
index dc76ebebc..49e138579 100644
--- a/icefall/profiler.py
+++ b/icefall/profiler.py
@@ -70,25 +70,17 @@ class FlopsProfiler(object):
                 module_flop_count.append([])
 
             if not hasattr(module, "__pre_hook_handle__"):
-                module.__pre_hook_handle__ = module.register_forward_pre_hook(
-                    pre_hook
-                )
+                module.__pre_hook_handle__ = module.register_forward_pre_hook(pre_hook)
 
             def post_hook(module, input, output):
                 if module_flop_count:
-                    module.__flops__ += sum(
-                        [elem[1] for elem in module_flop_count[-1]]
-                    )
+                    module.__flops__ += sum([elem[1] for elem in module_flop_count[-1]])
                     module_flop_count.pop()
 
             if not hasattr(module, "__post_hook_handle__"):
-                module.__post_hook_handle__ = module.register_forward_hook(
-                    post_hook
-                )
+                module.__post_hook_handle__ = module.register_forward_hook(post_hook)
 
-        self.model.apply(
-            partial(register_module_hooks, ignore_list=ignore_list)
-        )
+        self.model.apply(partial(register_module_hooks, ignore_list=ignore_list))
         self.started = True
         self.func_patched = True
 
@@ -194,9 +186,7 @@ def _prelu_flops_compute(input: Tensor, weight: Tensor):
     return input.numel()
 
 
-def _elu_flops_compute(
-    input: Tensor, alpha: float = 1.0, inplace: bool = False
-):
+def _elu_flops_compute(input: Tensor, alpha: float = 1.0, inplace: bool = False):
     return input.numel()
 
 
@@ -259,9 +249,7 @@ def _conv_flops_compute(
         output_dims.append(output_dim)
 
     filters_per_channel = out_channels // groups
-    conv_per_position_macs = (
-        int(_prod(kernel_dims)) * in_channels * filters_per_channel
-    )
+    conv_per_position_macs = int(_prod(kernel_dims)) * in_channels * filters_per_channel
     active_elements_count = batch_size * int(_prod(output_dims))
     overall_conv_macs = conv_per_position_macs * active_elements_count
     overall_conv_flops = 2 * overall_conv_macs
@@ -297,7 +285,6 @@ def _conv_trans_flops_compute(
 
     output_dims = []
     for idx, input_dim in enumerate(input_dims):
-
         output_dim = (
             input_dim
             + 2 * paddings[idx]
@@ -310,9 +297,7 @@ def _conv_trans_flops_compute(
     dilations = dilation if type(dilation) is tuple else (dilation, dilation)
 
     filters_per_channel = out_channels // groups
-    conv_per_position_macs = (
-        int(_prod(kernel_dims)) * in_channels * filters_per_channel
-    )
+    conv_per_position_macs = int(_prod(kernel_dims)) * in_channels * filters_per_channel
     active_elements_count = batch_size * int(_prod(input_dims))
     overall_conv_macs = conv_per_position_macs * active_elements_count
     overall_conv_flops = 2 * overall_conv_macs
@@ -389,9 +374,7 @@ def _upsample_flops_compute(input, **kwargs):
         else:
             return int(size), 0
     scale_factor = kwargs.get("scale_factor", None)
-    assert (
-        scale_factor is not None
-    ), "either size or scale_factor should be defined"
+    assert scale_factor is not None, "either size or scale_factor should be defined"
     flops = input.numel()
     if isinstance(scale_factor, tuple) and len(scale_factor) == len(input):
         flops * int(_prod(scale_factor))
@@ -593,12 +576,8 @@ def _patch_functionals():
     F.embedding = wrapFunc(F.embedding, _embedding_flops_compute)
 
     # swoosh functions in k2
-    k2.swoosh_l_forward = wrapFunc(
-        k2.swoosh_l_forward, _k2_swoosh_flops_compute
-    )
-    k2.swoosh_r_forward = wrapFunc(
-        k2.swoosh_r_forward, _k2_swoosh_flops_compute
-    )
+    k2.swoosh_l_forward = wrapFunc(k2.swoosh_l_forward, _k2_swoosh_flops_compute)
+    k2.swoosh_r_forward = wrapFunc(k2.swoosh_r_forward, _k2_swoosh_flops_compute)
     k2.swoosh_l = wrapFunc(k2.swoosh_l, _k2_swoosh_flops_compute)
     k2.swoosh_r = wrapFunc(k2.swoosh_r, _k2_swoosh_flops_compute)
 
@@ -612,9 +591,7 @@ def _patch_tensor_methods():
     torch.Tensor.bmm = wrapFunc(torch.Tensor.bmm, _matmul_flops_compute)
 
     torch.addmm = wrapFunc(torch.addmm, _addmm_flops_compute)
-    torch.Tensor.addmm = wrapFunc(
-        torch.Tensor.addmm, _tensor_addmm_flops_compute
-    )
+    torch.Tensor.addmm = wrapFunc(torch.Tensor.addmm, _tensor_addmm_flops_compute)
 
     torch.mul = wrapFunc(torch.mul, _mul_flops_compute)
     torch.Tensor.mul = wrapFunc(torch.Tensor.mul, _mul_flops_compute)
@@ -631,14 +608,10 @@ def _patch_tensor_methods():
 
     torch.tanh = wrapFunc(torch.tanh, _tanh_flops_compute)
 
-    torch.Tensor.softmax = wrapFunc(
-        torch.Tensor.softmax, _softmax_flops_compute
-    )
+    torch.Tensor.softmax = wrapFunc(torch.Tensor.softmax, _softmax_flops_compute)
 
     torch.sigmoid = wrapFunc(torch.sigmoid, _sigmoid_flops_compute)
-    torch.Tensor.sigmoid = wrapFunc(
-        torch.Tensor.sigmoid, _sigmoid_flops_compute
-    )
+    torch.Tensor.sigmoid = wrapFunc(torch.Tensor.sigmoid, _sigmoid_flops_compute)
 
 
 def _reload_functionals():
@@ -732,15 +705,11 @@ def _rnn_flops(flops, rnn_module, w_ih, w_hh, input_size):
         flops += rnn_module.hidden_size * 4
         # two hadamard _product and add for C state
         flops += (
-            rnn_module.hidden_size
-            + rnn_module.hidden_size
-            + rnn_module.hidden_size
+            rnn_module.hidden_size + rnn_module.hidden_size + rnn_module.hidden_size
         )
         # final hadamard
         flops += (
-            rnn_module.hidden_size
-            + rnn_module.hidden_size
-            + rnn_module.hidden_size
+            rnn_module.hidden_size + rnn_module.hidden_size + rnn_module.hidden_size
         )
     return flops
 
diff --git a/icefall/rnn_lm/check-onnx-streaming.py b/icefall/rnn_lm/check-onnx-streaming.py
index d51a4b76b..28b908f82 100755
--- a/icefall/rnn_lm/check-onnx-streaming.py
+++ b/icefall/rnn_lm/check-onnx-streaming.py
@@ -112,7 +112,6 @@ def main():
         for torch_v, onnx_v in zip(
             (torch_log_prob, torch_h0, torch_c0), (onnx_log_prob, onnx_h0, onnx_c0)
         ):
-
             assert torch.allclose(torch_v, onnx_v, atol=1e-5), (
                 torch_v.shape,
                 onnx_v.shape,
diff --git a/icefall/rnn_lm/train.py b/icefall/rnn_lm/train.py
index 3d206d139..0178b80bf 100755
--- a/icefall/rnn_lm/train.py
+++ b/icefall/rnn_lm/train.py
@@ -463,7 +463,6 @@ def train_one_epoch(
     cur_batch_idx = params.get("cur_batch_idx", 0)
 
     for batch_idx, batch in enumerate(train_dl):
-
         if batch_idx < cur_batch_idx:
             continue
         cur_batch_idx = batch_idx
diff --git a/icefall/shared/make_kn_lm.py b/icefall/shared/make_kn_lm.py
index 7150297d6..231aca7f1 100755
--- a/icefall/shared/make_kn_lm.py
+++ b/icefall/shared/make_kn_lm.py
@@ -225,7 +225,6 @@ class NgramCounts:
         for n in range(0, self.ngram_order - 1):
             this_order_counts = self.counts[n]
             for hist, counts_for_hist in this_order_counts.items():
-
                 n_star_star = 0
                 for w in counts_for_hist.word_to_count.keys():
                     n_star_star += len(counts_for_hist.word_to_context[w])
@@ -424,7 +423,6 @@ class NgramCounts:
 
 
 if __name__ == "__main__":
-
     ngram_counts = NgramCounts(args.ngram_order)
 
     if args.text is None:
diff --git a/icefall/transformer_lm/model.py b/icefall/transformer_lm/model.py
index 79dda3168..c78cf1821 100644
--- a/icefall/transformer_lm/model.py
+++ b/icefall/transformer_lm/model.py
@@ -103,7 +103,6 @@ class TransformerLM(torch.nn.Module):
         return nll_loss
 
     def score_token(self, x: torch.Tensor, x_lens: torch.Tensor, state=None):
-
         bs = x.size(0)
 
         state = None
diff --git a/icefall/utils.py b/icefall/utils.py
index 947d79438..a9e8a81b9 100644
--- a/icefall/utils.py
+++ b/icefall/utils.py
@@ -263,6 +263,70 @@ def get_texts(
         return aux_labels.tolist()
 
 
+def encode_supervisions_otc(
+    supervisions: dict,
+    subsampling_factor: int,
+    token_ids: Optional[List[List[int]]] = None,
+) -> Tuple[torch.Tensor, Union[List[str], List[List[int]]]]:
+    """
+    Encodes Lhotse's ``batch["supervisions"]`` dict into
+    a pair of torch Tensor, and a list of transcription strings or token indexes
+
+    The supervision tensor has shape ``(batch_size, 3)``.
+    Its second dimension contains information about sequence index [0],
+    start frames [1] and num frames [2].
+
+    The batch items might become re-ordered during this operation -- the
+    returned tensor and list of strings are guaranteed to be consistent with
+    each other.
+    """
+    supervision_segments = torch.stack(
+        (
+            supervisions["sequence_idx"],
+            torch.div(
+                supervisions["start_frame"],
+                subsampling_factor,
+                rounding_mode="floor",
+            ),
+            torch.div(
+                supervisions["num_frames"],
+                subsampling_factor,
+                rounding_mode="floor",
+            ),
+        ),
+        1,
+    ).to(torch.int32)
+
+    indices = torch.argsort(supervision_segments[:, 2], descending=True)
+    supervision_segments = supervision_segments[indices]
+
+    ids = []
+    verbatim_texts = []
+    sorted_ids = []
+    sorted_verbatim_texts = []
+
+    for cut in supervisions["cut"]:
+        id = cut.id
+        if hasattr(cut.supervisions[0], "verbatim_text"):
+            verbatim_text = cut.supervisions[0].verbatim_text
+        else:
+            verbatim_text = ""
+        ids.append(id)
+        verbatim_texts.append(verbatim_text)
+
+    for index in indices.tolist():
+        sorted_ids.append(ids[index])
+        sorted_verbatim_texts.append(verbatim_texts[index])
+
+    if token_ids is None:
+        texts = supervisions["text"]
+        res = [texts[idx] for idx in indices]
+    else:
+        res = [token_ids[idx] for idx in indices]
+
+    return supervision_segments, res, sorted_ids, sorted_verbatim_texts
+
+
 @dataclass
 class DecodingResults:
     # timestamps[i][k] contains the frame number on which tokens[i][k]
@@ -419,7 +483,7 @@ def load_alignments(filename: str) -> Tuple[int, Dict[str, List[int]]]:
 
 
 def store_transcripts(
-    filename: Pathlike, texts: Iterable[Tuple[str, str, str]]
+    filename: Pathlike, texts: Iterable[Tuple[str, str, str]], char_level: bool = False
 ) -> None:
     """Save predicted results and reference transcripts to a file.
 
@@ -434,8 +498,11 @@ def store_transcripts(
     Returns:
       Return None.
     """
-    with open(filename, "w") as f:
+    with open(filename, "w", encoding="utf8") as f:
         for cut_id, ref, hyp in texts:
+            if char_level:
+                ref = list("".join(ref))
+                hyp = list("".join(hyp))
             print(f"{cut_id}:\tref={ref}", file=f)
             print(f"{cut_id}:\thyp={hyp}", file=f)
 
@@ -456,7 +523,7 @@ def store_transcripts_and_timestamps(
     Returns:
       Return None.
     """
-    with open(filename, "w") as f:
+    with open(filename, "w", encoding="utf8") as f:
         for cut_id, ref, hyp, time_ref, time_hyp in texts:
             print(f"{cut_id}:\tref={ref}", file=f)
             print(f"{cut_id}:\thyp={hyp}", file=f)
@@ -493,6 +560,7 @@ def write_error_stats(
     test_set_name: str,
     results: List[Tuple[str, str]],
     enable_log: bool = True,
+    compute_CER: bool = False,
     sclite_mode: bool = False,
 ) -> float:
     """Write statistics based on predicted results and reference transcripts.
@@ -521,7 +589,7 @@ def write_error_stats(
           The reference word `SIR` is missing in the predicted
           results (a deletion error).
       results:
-        An iterable of tuples. The first element is the cur_id, the second is
+        An iterable of tuples. The first element is the cut_id, the second is
         the reference transcript and the third element is the predicted result.
       enable_log:
         If True, also print detailed WER to the console.
@@ -538,6 +606,14 @@ def write_error_stats(
     words: Dict[str, List[int]] = defaultdict(lambda: [0, 0, 0, 0, 0])
     num_corr = 0
     ERR = "*"
+
+    if compute_CER:
+        for i, res in enumerate(results):
+            cut_id, ref, hyp = res
+            ref = list("".join(ref))
+            hyp = list("".join(hyp))
+            results[i] = (cut_id, ref, hyp)
+
     for cut_id, ref, hyp in results:
         ali = kaldialign.align(ref, hyp, ERR, sclite_mode=sclite_mode)
         for ref_word, hyp_word in ali:
@@ -1362,13 +1438,16 @@ def measure_gradient_norms(model: nn.Module, norm: str = "l1") -> Dict[str, floa
 
 
 def get_parameter_groups_with_lrs(
-    model: nn.Module, lr: float, include_names: bool = False
+    model: nn.Module,
+    lr: float,
+    include_names: bool = False,
+    freeze_modules: List[str] = [],
 ) -> List[dict]:
     """
     This is for use with the ScaledAdam optimizers (more recent versions that accept lists of
     named-parameters; we can, if needed, create a version without the names).
 
-    It provides a way to specifiy learning-rate scales inside the module, so that if
+    It provides a way to specify learning-rate scales inside the module, so that if
     any nn.Module in the hierarchy has a floating-point parameter 'lr_scale', it will
     scale the LR of any parameters inside that module or its submodules.  Note: you
     can set module parameters outside the __init__ function, e.g.:
@@ -1386,6 +1465,8 @@ def get_parameter_groups_with_lrs(
          ...   ]
 
     """
+    named_modules = list(model.named_modules())
+
     # flat_lr_scale just contains the lr_scale explicitly specified
     # for each prefix of the name, e.g. 'encoder.layers.3', these need
     # to be multiplied for all prefix of the name of any given parameter.
@@ -1405,6 +1486,15 @@ def get_parameter_groups_with_lrs(
         split_name = name.split(".")
         # caution: as a special case, if the name is '', split_name will be [ '' ].
         prefix = split_name[0]
+        if prefix == "module":  # DDP
+            module_name = split_name[1]
+            if module_name in freeze_modules:
+                logging.info(f"Remove {name} from parameters")
+                continue
+        else:
+            if prefix in freeze_modules:
+                logging.info(f"Remove {name} from parameters")
+                continue
         cur_lr = lr * flat_lr_scale[prefix]
         if prefix != "":
             cur_lr *= flat_lr_scale[""]
@@ -1517,10 +1607,10 @@ def tokenize_by_bpe_model(
     chars = pattern.split(txt.upper())
     mix_chars = [w for w in chars if len(w.strip()) > 0]
     for ch_or_w in mix_chars:
-        # ch_or_w is a single CJK charater(i.e., "你"), do nothing.
+        # ch_or_w is a single CJK character(i.e., "你"), do nothing.
         if pattern.fullmatch(ch_or_w) is not None:
             tokens.append(ch_or_w)
-        # ch_or_w contains non-CJK charaters(i.e., " IT'S OKAY "),
+        # ch_or_w contains non-CJK characters(i.e., " IT'S OKAY "),
         # encode ch_or_w using bpe_model.
         else:
             for p in sp.encode_as_pieces(ch_or_w):
@@ -1534,7 +1624,7 @@ def tokenize_by_CJK_char(line: str) -> str:
     """
     Tokenize a line of text with CJK char.
 
-    Note: All return charaters will be upper case.
+    Note: All return characters will be upper case.
 
     Example:
       input = "你好世界是 hello world 的中文"
@@ -1827,7 +1917,7 @@ def parse_bpe_timestamps_and_texts(
         A k2.Fsa with best_paths.arcs.num_axes() == 3, i.e.
         containing multiple FSAs, which is expected to be the result
         of k2.shortest_path (otherwise the returned values won't
-        be meaningful). Its attribtutes `labels` and `aux_labels`
+        be meaningful). Its attributes `labels` and `aux_labels`
         are both BPE tokens.
       sp:
         The BPE model.
@@ -1887,7 +1977,7 @@ def parse_timestamps_and_texts(
         A k2.Fsa with best_paths.arcs.num_axes() == 3, i.e.
         containing multiple FSAs, which is expected to be the result
         of k2.shortest_path (otherwise the returned values won't
-        be meaningful). Attribtute `labels` is the prediction unit,
+        be meaningful). Attribute `labels` is the prediction unit,
         e.g., phone or BPE tokens. Attribute `aux_labels` is the word index.
       word_table:
         The word symbol table.
@@ -1955,7 +2045,7 @@ def parse_fsa_timestamps_and_texts(
 ) -> Tuple[List[Tuple[float, float]], List[List[str]]]:
     """Parse timestamps (in seconds) and texts for given decoded fsa paths.
     Currently it supports two cases:
-    (1) ctc-decoding, the attribtutes `labels` and `aux_labels`
+    (1) ctc-decoding, the attributes `labels` and `aux_labels`
         are both BPE tokens. In this case, sp should be provided.
     (2) HLG-based 1best, the attribtute `labels` is the prediction unit,
         e.g., phone or BPE tokens; attribute `aux_labels` is the word index.
diff --git a/requirements-ci.txt b/requirements-ci.txt
index 2433e190b..1eba69764 100644
--- a/requirements-ci.txt
+++ b/requirements-ci.txt
@@ -20,9 +20,11 @@ kaldialign==0.7.1
 sentencepiece==0.1.96
 tensorboard==2.8.0
 typeguard==2.13.3
+black==22.3.0
 multi_quantization
 
 onnx
 onnxmltools
 onnxruntime
 kaldifst
+kaldi-decoder
diff --git a/requirements.txt b/requirements.txt
index a07f6b7c7..5a8326619 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,7 +1,9 @@
 kaldifst
 kaldilm
 kaldialign
+kaldi-decoder
 sentencepiece>=0.1.96
 tensorboard
 typeguard
 dill
+black==22.3.0