Merge branch 'master' into cr-ctc

.github/scripts/docker/generate_build_matrix.py

@@ -45,15 +45,17 @@ def get_torchaudio_version(torch_version):
 def get_matrix():
     k2_version = "1.24.4.dev20240223"
     kaldifeat_version = "1.25.4.dev20240223"
-    version = "20240725"
+    version = "20240905"
     python_version = ["3.8", "3.9", "3.10", "3.11", "3.12"]
     torch_version = []
-    torch_version += ["1.13.0", "1.13.1"]
+    #  torch_version += ["1.13.0", "1.13.1"]
-    torch_version += ["2.0.0", "2.0.1"]
+    #  torch_version += ["2.0.0", "2.0.1"]
-    torch_version += ["2.1.0", "2.1.1", "2.1.2"]
+    #  torch_version += ["2.1.0", "2.1.1", "2.1.2"]
-    torch_version += ["2.2.0", "2.2.1", "2.2.2"]
+    #  torch_version += ["2.2.0", "2.2.1", "2.2.2"]
+    # Test only torch >= 2.3.0
     torch_version += ["2.3.0", "2.3.1"]
     torch_version += ["2.4.0"]
+    torch_version += ["2.4.1"]
 
     matrix = []
     for p in python_version:
@@ -82,6 +84,9 @@ def get_matrix():
             elif t == "2.4.0":
                 k2_version_2 = "1.24.4.dev20240725"
                 kaldifeat_version_2 = "1.25.4.dev20240725"
+            elif t == "2.4.1":
+                k2_version_2 = "1.24.4.dev20240905"
+                kaldifeat_version_2 = "1.25.4.dev20240905"
 
             matrix.append(
                 {

.github/scripts/multi-zh-hans.sh

@@ -16,6 +16,48 @@ log "pwd: $PWD"
 
 cd egs/multi_zh-hans/ASR
 
+repo_url=https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+pushd $repo
+cd exp
+git lfs pull --include pretrained.pt
+ln -s pretrained.pt epoch-99.pt
+cd ../data/lang_bpe_2000
+ls -lh
+git lfs pull --include L.pt L_disambig.pt Linv.pt bpe.model
+git lfs pull --include "*.model"
+ls -lh
+popd
+
+log "--------------------------------------------"
+log "Export non-streaming ONNX transducer models "
+log "--------------------------------------------"
+./zipformer/export-onnx.py \
+  --tokens $repo/data/lang_bpe_2000/tokens.txt \
+  --use-averaged-model 0 \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --causal False
+
+ls -lh $repo/exp
+
+./zipformer/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_2000/tokens.txt \
+  $repo/test_wavs/DEV_T0000000000.wav \
+  $repo/test_wavs/DEV_T0000000001.wav \
+  $repo/test_wavs/DEV_T0000000002.wav \
+  $repo/test_wavs/TEST_MEETING_T0000000113.wav \
+  $repo/test_wavs/TEST_MEETING_T0000000219.wav \
+  $repo/test_wavs/TEST_MEETING_T0000000351.wav
+
+rm -rf $repo
+
 repo_url=https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-ctc-streaming-2023-11-05
 log "Downloading pre-trained model from $repo_url"
 GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url

.github/scripts/run-gigaspeech-pruned-transducer-stateless2-2022-05-12.sh

@@ -29,8 +29,8 @@ if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" ==
   ls -lh data/fbank
   ls -lh pruned_transducer_stateless2/exp
 
-  ln -s data/fbank/cuts_DEV.jsonl.gz data/fbank/gigaspeech_cuts_DEV.jsonl.gz
+  ln -sf data/fbank/cuts_DEV.jsonl.gz data/fbank/gigaspeech_cuts_DEV.jsonl.gz
-  ln -s data/fbank/cuts_TEST.jsonl.gz data/fbank/gigaspeech_cuts_TEST.jsonl.gz
+  ln -sf data/fbank/cuts_TEST.jsonl.gz data/fbank/gigaspeech_cuts_TEST.jsonl.gz
 
   log "Decoding dev and test"
 

.github/scripts/test-onnx-export.sh

@@ -25,6 +25,7 @@ popd
 
 log "Export via torch.jit.script()"
 ./zipformer/export.py \
+  --use-averaged-model 0 \
   --exp-dir $repo/exp \
   --tokens $repo/data/lang_bpe_500/tokens.txt \
   --epoch 99 \

.github/workflows/build-doc.yml

@@ -26,6 +26,8 @@ on:
   pull_request:
     types: [labeled]
 
+  workflow_dispatch:
+
 concurrency:
   group: build_doc-${{ github.ref }}
   cancel-in-progress: true

.github/workflows/build-docker-image.yml

@@ -16,7 +16,7 @@ jobs:
       fail-fast: false
       matrix:
         os: [ubuntu-latest]
-        image: ["torch2.4.0-cuda12.4", "torch2.4.0-cuda12.1", "torch2.4.0-cuda11.8", "torch2.3.1-cuda12.1", "torch2.3.1-cuda11.8", "torch2.2.2-cuda12.1", "torch2.2.2-cuda11.8", "torch2.2.1-cuda12.1", "torch2.2.1-cuda11.8", "torch2.2.0-cuda12.1", "torch2.2.0-cuda11.8", "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"]
+        image: ["torch2.4.1-cuda12.4", "torch2.4.1-cuda12.1", "torch2.4.1-cuda11.8", "torch2.4.0-cuda12.4", "torch2.4.0-cuda12.1", "torch2.4.0-cuda11.8", "torch2.3.1-cuda12.1", "torch2.3.1-cuda11.8", "torch2.2.2-cuda12.1", "torch2.2.2-cuda11.8", "torch2.2.1-cuda12.1", "torch2.2.1-cuda11.8", "torch2.2.0-cuda12.1", "torch2.2.0-cuda11.8", "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

.github/workflows/run-gigaspeech-2022-05-13.yml

@@ -33,6 +33,8 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+  workflow_dispatch:
+
 concurrency:
   group: run_gigaspeech_2022_05_13-${{ github.ref }}
   cancel-in-progress: true
@@ -119,7 +121,7 @@ jobs:
           find exp/greedy_search -name "log-*" -exec grep -n --color "best for test" {} + | sort -n -k2
 
       - name: Upload decoding results for gigaspeech pruned_transducer_stateless2
-        uses: actions/upload-artifact@v2
+        uses: actions/upload-artifact@v4
         if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
         with:
           name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-gigaspeech-pruned_transducer_stateless2-2022-05-12

.github/workflows/run-gigaspeech-zipformer-2023-10-17.yml

@@ -42,7 +42,7 @@ concurrency:
 
 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'
+    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' || github.event_name == 'workflow_dispatch'
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
@@ -133,7 +133,7 @@ jobs:
           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
+        uses: actions/upload-artifact@v4
         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

.github/workflows/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml

@@ -16,6 +16,8 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+  workflow_dispatch:
+
 concurrency:
   group: run_librispeech_lstm_transducer_stateless2_2022_09_03-${{ github.ref }}
   cancel-in-progress: true
@@ -156,7 +158,7 @@ jobs:
           find modified_beam_search_LODR  -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
 
       - name: Upload decoding results for lstm_transducer_stateless2
-        uses: actions/upload-artifact@v2
+        uses: actions/upload-artifact@v4
         if: github.event_name == 'schedule' || github.event.label.name == 'shallow-fusion' || github.event.label.name == 'LODR'
         with:
           name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu-lstm_transducer_stateless2-2022-09-03

.github/workflows/run-multi-corpora-zipformer.yml

@@ -23,6 +23,8 @@ on:
   pull_request:
     types: [labeled]
 
+  workflow_dispatch:
+
 concurrency:
   group: run_multi-corpora_zipformer-${{ github.ref }}
   cancel-in-progress: true

.github/workflows/run-ptb-rnn-lm.yml

@@ -16,6 +16,8 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+  workflow_dispatch:
+
 concurrency:
   group: run_ptb_rnn_lm_training-${{ github.ref }}
   cancel-in-progress: true
@@ -64,7 +66,7 @@ jobs:
           ./train-rnn-lm.sh --world-size 1 --num-epochs 5 --use-epoch 4 --use-avg 2
 
       - name: Upload pretrained models
-        uses: actions/upload-artifact@v2
+        uses: actions/upload-artifact@v4
         if: github.event.label.name == 'ready' || github.event.label.name == 'rnnlm' || github.event_name == 'push' || github.event_name == 'schedule'
         with:
           name: python-${{ matrix.python-version }}-ubuntu-rnn-lm-ptb

.github/workflows/run-swbd-conformer-ctc.yml

@@ -23,6 +23,8 @@ on:
   pull_request:
     types: [labeled]
 
+  workflow_dispatch:
+
 concurrency:
   group: run-swbd-conformer_ctc-${{ github.ref }}
   cancel-in-progress: true

.github/workflows/run-wenetspeech-pruned-transducer-stateless2.yml

@@ -23,6 +23,8 @@ on:
   pull_request:
     types: [labeled]
 
+  workflow_dispatch:
+
 concurrency:
   group: run_wenetspeech_pruned_transducer_stateless2-${{ github.ref }}
   cancel-in-progress: true

.github/workflows/style_check.yml

@@ -24,6 +24,8 @@ on:
     branches:
       - master
 
+  workflow_dispatch:
+
 concurrency:
   group: style_check-${{ github.ref }}
   cancel-in-progress: true

.github/workflows/test-ncnn-export.yml

@@ -16,6 +16,8 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+  workflow_dispatch:
+
 concurrency:
   group: test_ncnn_export-${{ github.ref }}
   cancel-in-progress: true

.github/workflows/test-onnx-export.yml

@@ -16,6 +16,8 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+  workflow_dispatch:
+
 concurrency:
   group: test_onnx_export-${{ github.ref }}
   cancel-in-progress: true

.github/workflows/test.yml

@@ -105,7 +105,7 @@ jobs:
               cd ../zipformer
               pytest -v -s
 
-      - uses: actions/upload-artifact@v2
+      - uses: actions/upload-artifact@v4
         with:
           path: egs/librispeech/ASR/zipformer/swoosh.pdf
-          name: swoosh.pdf
+          name: swoosh-${{ matrix.python-version }}-${{ matrix.torch-version }}

docker/torch2.4.1-cuda11.8.dockerfile

@@ -0,0 +1,73 @@
+FROM pytorch/pytorch:2.4.1-cuda11.8-cudnn9-devel
+# python 3.10
+
+ENV LC_ALL C.UTF-8
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# python 3.10
+ARG K2_VERSION="1.24.4.dev20240905+cuda11.8.torch2.4.1"
+ARG KALDIFEAT_VERSION="1.25.4.dev20240905+cuda11.8.torch2.4.1"
+ARG TORCHAUDIO_VERSION="2.4.1+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==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torchaudio/ \
+      k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \
+      git+https://github.com/lhotse-speech/lhotse \
+      kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \
+      kaldi_native_io \
+      kaldialign \
+      kaldifst \
+      kaldilm \
+      sentencepiece>=0.1.96 \
+      tensorboard \
+      typeguard \
+      dill \
+      onnx \
+      onnxruntime \
+      onnxmltools \
+      onnxoptimizer \
+      onnxsim \
+      multi_quantization \
+      typeguard \
+      numpy \
+      pytest \
+      graphviz
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+    cd /workspace/icefall && \
+    pip install --no-cache-dir -r requirements.txt
+
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+WORKDIR /workspace/icefall

docker/torch2.4.1-cuda12.1.dockerfile

@@ -0,0 +1,73 @@
+FROM pytorch/pytorch:2.4.1-cuda12.1-cudnn9-devel
+# python 3.10
+
+ENV LC_ALL C.UTF-8
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# python 3.10
+ARG K2_VERSION="1.24.4.dev20240905+cuda12.1.torch2.4.1"
+ARG KALDIFEAT_VERSION="1.25.4.dev20240905+cuda12.1.torch2.4.1"
+ARG TORCHAUDIO_VERSION="2.4.1+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==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torchaudio/ \
+      k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \
+      git+https://github.com/lhotse-speech/lhotse \
+      kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \
+      kaldi_native_io \
+      kaldialign \
+      kaldifst \
+      kaldilm \
+      sentencepiece>=0.1.96 \
+      tensorboard \
+      typeguard \
+      dill \
+      onnx \
+      onnxruntime \
+      onnxmltools \
+      onnxoptimizer \
+      onnxsim \
+      multi_quantization \
+      typeguard \
+      numpy \
+      pytest \
+      graphviz
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+    cd /workspace/icefall && \
+    pip install --no-cache-dir -r requirements.txt
+
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+WORKDIR /workspace/icefall

docker/torch2.4.1-cuda12.4.dockerfile

@@ -0,0 +1,73 @@
+FROM pytorch/pytorch:2.4.1-cuda12.4-cudnn9-devel
+# python 3.10
+
+ENV LC_ALL C.UTF-8
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# python 3.10
+ARG K2_VERSION="1.24.4.dev20240905+cuda12.4.torch2.4.1"
+ARG KALDIFEAT_VERSION="1.25.4.dev20240905+cuda12.4.torch2.4.1"
+ARG TORCHAUDIO_VERSION="2.4.1+cu124"
+
+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==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torchaudio/ \
+      k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \
+      git+https://github.com/lhotse-speech/lhotse \
+      kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \
+      kaldi_native_io \
+      kaldialign \
+      kaldifst \
+      kaldilm \
+      sentencepiece>=0.1.96 \
+      tensorboard \
+      typeguard \
+      dill \
+      onnx \
+      onnxruntime \
+      onnxmltools \
+      onnxoptimizer \
+      onnxsim \
+      multi_quantization \
+      typeguard \
+      numpy \
+      pytest \
+      graphviz
+
+RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
+    cd /workspace/icefall && \
+    pip install --no-cache-dir -r requirements.txt
+
+ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
+
+WORKDIR /workspace/icefall

docs/source/docker/intro.rst

@@ -34,6 +34,12 @@ which will give you something like below:
 
 .. code-block:: bash
 
+  "torch2.4.1-cuda12.4"
+  "torch2.4.1-cuda12.1"
+  "torch2.4.1-cuda11.8"
+  "torch2.4.0-cuda12.4"
+  "torch2.4.0-cuda12.1"
+  "torch2.4.0-cuda11.8"
   "torch2.3.1-cuda12.1"
   "torch2.3.1-cuda11.8"
   "torch2.2.2-cuda12.1"

egs/alimeeting/ASR/prepare.sh

@@ -87,7 +87,7 @@ fi
 if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
   log "Stage 3: Prepare musan manifest"
   # We assume that you have downloaded the musan corpus
-  # to data/musan
+  # to $dl_dir/musan
   if [ ! -f data/manifests/.musan_manifests.done ]; then
     log "It may take 6 minutes"
     mkdir -p data/manifests

egs/alimeeting/ASR_v2/local/prepare_alimeeting_gss.sh

@@ -58,7 +58,7 @@ if [ $stage -le 4 ]; then
   # for train, we use smaller context and larger batches to speed-up processing
   for JOB in $(seq $nj); do
     gss enhance cuts $EXP_DIR/cuts_train.jsonl.gz \
-      $EXP_DIR/cuts_per_segment_train_split$nj/cuts_per_segment_train.JOB.jsonl.gz $EXP_DIR/enhanced \
+      $EXP_DIR/cuts_per_segment_train_split$nj/cuts_per_segment_train.$JOB.jsonl.gz $EXP_DIR/enhanced \
       --bss-iterations 10 \
       --context-duration 5.0 \
       --use-garbage-class \
@@ -77,7 +77,7 @@ if [ $stage -le 5 ]; then
   for part in eval test; do
     for JOB in $(seq $nj); do
       gss enhance cuts $EXP_DIR/cuts_${part}.jsonl.gz \
-      $EXP_DIR/cuts_per_segment_${part}_split$nj/cuts_per_segment_${part}.JOB.jsonl.gz \
+      $EXP_DIR/cuts_per_segment_${part}_split$nj/cuts_per_segment_${part}.$JOB.jsonl.gz \
       $EXP_DIR/enhanced \
       --bss-iterations 10 \
       --context-duration 15.0 \

egs/alimeeting/ASR_v2/prepare.sh

@@ -65,7 +65,7 @@ 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 data/musan
+  # to $dl_dir/musan
   mkdir -p data/manifests
   lhotse prepare musan $dl_dir/musan data/manifests
 fi

egs/alimeeting/ASR_v2/pruned_transducer_stateless7/asr_datamodule.py

@@ -82,7 +82,7 @@ class AlimeetingAsrDataModule:
         group.add_argument(
             "--manifest-dir",
             type=Path,
-            default=Path("data/manifests"),
+            default=Path("data/fbank"),
             help="Path to directory with train/valid/test cuts.",
         )
         group.add_argument(
@@ -327,9 +327,11 @@ class AlimeetingAsrDataModule:
     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)))
+            input_strategy=(
-            if self.args.on_the_fly_feats
+                OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
-            else PrecomputedFeatures(),
+                if self.args.on_the_fly_feats
+                else PrecomputedFeatures()
+            ),
             return_cuts=True,
         )
         sampler = DynamicBucketingSampler(

egs/ami/ASR/local/prepare_ami_gss.sh

@@ -58,7 +58,7 @@ if [ $stage -le 4 ]; then
   # for train, we use smaller context and larger batches to speed-up processing
   for JOB in $(seq $nj); do
     gss enhance cuts $EXP_DIR/cuts_train.jsonl.gz \
-      $EXP_DIR/cuts_per_segment_train_split$nj/cuts_per_segment_train.JOB.jsonl.gz $EXP_DIR/enhanced \
+      $EXP_DIR/cuts_per_segment_train_split$nj/cuts_per_segment_train.$JOB.jsonl.gz $EXP_DIR/enhanced \
       --bss-iterations 10 \
       --context-duration 5.0 \
       --use-garbage-class \
@@ -77,7 +77,7 @@ if [ $stage -le 5 ]; then
   for part in dev test; do
     for JOB in $(seq $nj); do
       gss enhance cuts $EXP_DIR/cuts_${part}.jsonl.gz \
-      $EXP_DIR/cuts_per_segment_${part}_split$nj/cuts_per_segment_${part}.JOB.jsonl.gz \
+      $EXP_DIR/cuts_per_segment_${part}_split$nj/cuts_per_segment_${part}.$JOB.jsonl.gz \
       $EXP_DIR/enhanced \
       --bss-iterations 10 \
       --context-duration 15.0 \

egs/librispeech/ASR/zipformer/export-onnx-streaming.py

@@ -74,7 +74,6 @@ import onnx
 import torch
 import torch.nn as nn
 from decoder import Decoder
-from onnxconverter_common import float16
 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
@@ -756,6 +755,7 @@ def main():
     logging.info(f"Exported joiner to {joiner_filename}")
 
     if(params.fp16) :
+        from onnxconverter_common import float16
         logging.info("Generate fp16 models")
 
         encoder = onnx.load(encoder_filename)

egs/librispeech/ASR/zipformer/zipformer.py

@@ -191,6 +191,7 @@ class Zipformer2(EncoderInterface):
                     dim=encoder_dim[i],
                     downsample=downsampling_factor[i],
                     dropout=dropout,
+                    causal=causal,
                 )
 
             encoders.append(encoder)
@@ -198,7 +199,10 @@ class Zipformer2(EncoderInterface):
         self.encoders = nn.ModuleList(encoders)
 
         self.downsample_output = SimpleDownsample(
-            max(encoder_dim), downsample=output_downsampling_factor, dropout=dropout
+            max(encoder_dim),
+            downsample=output_downsampling_factor,
+            dropout=dropout,
+            causal=causal,
         )
 
     def get_feature_masks(self, x: Tensor) -> Union[List[float], List[Tensor]]:
@@ -1217,11 +1221,16 @@ class DownsampledZipformer2Encoder(nn.Module):
     """
 
     def __init__(
-        self, encoder: nn.Module, dim: int, downsample: int, dropout: FloatLike
+        self,
+        encoder: nn.Module,
+        dim: int,
+        downsample: int,
+        dropout: FloatLike,
+        causal: bool,
     ):
         super(DownsampledZipformer2Encoder, self).__init__()
         self.downsample_factor = downsample
-        self.downsample = SimpleDownsample(dim, downsample, dropout)
+        self.downsample = SimpleDownsample(dim, downsample, dropout, causal)
         self.num_layers = encoder.num_layers
         self.encoder = encoder
         self.upsample = SimpleUpsample(dim, downsample)
@@ -1310,9 +1319,12 @@ class SimpleDownsample(torch.nn.Module):
     Does downsampling with attention, by weighted sum, and a projection..
     """
 
-    def __init__(self, channels: int, downsample: int, dropout: FloatLike):
+    def __init__(
+        self, channels: int, downsample: int, dropout: FloatLike, causal: bool
+    ):
         super(SimpleDownsample, self).__init__()
 
+        self.causal = causal
         self.bias = nn.Parameter(torch.zeros(downsample))
 
         self.name = None  # will be set from training code
@@ -1333,9 +1345,18 @@ class SimpleDownsample(torch.nn.Module):
         # Pad to an exact multiple of self.downsample
         # 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)
+        if self.causal and torch.jit.is_tracing():
-        assert src.shape[0] == d_seq_len * ds
+            assert (
+                pad == 0
+            ), f"pad should be zero for exporting streaming models. Given {pad}"
+
+        # If we are exporting a streaming model, then we skip the if statement
+        if not self.causal or not torch.jit.is_tracing():
+            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.shape, d_seq_len, ds)
 
         src = src.reshape(d_seq_len, ds, batch_size, in_channels)
 
@@ -1609,7 +1630,11 @@ class RelPositionMultiheadAttentionWeights(nn.Module):
         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, (p.shape[-1], num_heads, pos_head_dim)
+        assert p.shape[-1] == num_heads * pos_head_dim, (
+            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.

egs/wenetspeech/KWS/prepare.sh

@@ -63,8 +63,8 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
     ln -svf $(realpath ./open-commands/CN/small/commands.txt) commands_small.txt
     ln -svf $(realpath ./open-commands/CN/large/commands.txt) commands_large.txt
     pushd open-commands
-    ./script/prepare.sh --stage 1 --stop-stage 1
+    ./scripts/prepare.sh --stage 1 --stop-stage 1
-    ./script/prepare.sh --stage 3 --stop-stage 5
+    ./scripts/prepare.sh --stage 3 --stop-stage 5
     popd
     popd
     pushd data/fbank