Merge remote-tracking branch 'k2-fsa/master'

.flake8

@@ -1,7 +1,7 @@
 [flake8]
 show-source=true
 statistics=true
-max-line-length = 80
+max-line-length = 88
 per-file-ignores =
     # line too long
     icefall/diagnostics.py: E501,
@@ -11,7 +11,8 @@ per-file-ignores =
     egs/*/ASR/*/scaling.py: E501,
     egs/librispeech/ASR/lstm_transducer_stateless*/*.py: E501, E203
     egs/librispeech/ASR/conv_emformer_transducer_stateless*/*.py: E501, E203
-    egs/librispeech/ASR/conformer_ctc2/*py: E501,
+    egs/librispeech/ASR/conformer_ctc*/*py: E501,
+    egs/librispeech/ASR/zipformer_mmi/*.py: E501, E203
     egs/librispeech/ASR/RESULTS.md: E999,
 
     # invalid escape sequence (cause by tex formular), W605

.git-blame-ignore-revs

@@ -0,0 +1,3 @@
+# Migrate to 88 characters per line (see: https://github.com/lhotse-speech/lhotse/issues/890)
+107df3b115a58f1b68a6458c3f94a130004be34c
+d31db010371a4128856480382876acdc0d1739ed

.github/scripts/run-librispeech-conformer-ctc3-2022-11-28.sh

@@ -0,0 +1,123 @@
+#!/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://huggingface.co/Zengwei/icefall-asr-librispeech-conformer-ctc3-2022-11-27
+
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+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/LG.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/lm/G_4_gram.pt"
+git lfs pull --include "exp/jit_trace.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Decode with models exported by torch.jit.trace()"
+
+for m in ctc-decoding 1best; do
+  ./conformer_ctc3/jit_pretrained.py \
+    --model-filename $repo/exp/jit_trace.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --G $repo/data/lm/G_4_gram.pt \
+    --method $m \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+log "Export to torchscript model"
+
+./conformer_ctc3/export.py \
+  --exp-dir $repo/exp \
+  --lang-dir $repo/data/lang_bpe_500 \
+  --jit-trace 1 \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.trace()"
+
+for m in ctc-decoding 1best; do
+  ./conformer_ctc3/jit_pretrained.py \
+    --model-filename $repo/exp/jit_trace.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --G $repo/data/lm/G_4_gram.pt \
+    --method $m \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for m in ctc-decoding 1best; do
+  ./conformer_ctc3/pretrained.py \
+    --checkpoint $repo/exp/pretrained.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --G $repo/data/lm/G_4_gram.pt \
+    --method $m \
+    --sample-rate 16000 \
+    $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 conformer_ctc3/exp
+  ln -s $PWD/$repo/exp/pretrained.pt conformer_ctc3/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh conformer_ctc3/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in ctc-decoding 1best; do
+    log "Decoding with $method"
+    ./conformer_ctc3/decode.py \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --exp-dir conformer_ctc3/exp/ \
+      --max-duration $max_duration \
+      --decoding-method $method \
+      --lm-dir data/lm
+  done
+
+  rm conformer_ctc3/exp/*.pt
+fi

.github/scripts/run-librispeech-conv-emformer-transducer-stateless2-2022-12-05.sh

@@ -0,0 +1,79 @@
+#!/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://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05
+
+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-epoch-30-avg-10-averaged.pt"
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+cd exp
+ln -s pretrained-epoch-30-avg-10-averaged.pt epoch-99.pt
+popd
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+log  "Install ncnn and pnnx"
+
+# We are using a modified ncnn here. Will try to merge it to the official repo
+# of ncnn
+git clone https://github.com/csukuangfj/ncnn
+pushd ncnn
+git submodule init
+git submodule update python/pybind11
+python3 setup.py bdist_wheel
+ls -lh dist/
+pip install dist/*.whl
+cd tools/pnnx
+mkdir build
+cd build
+cmake -D Python3_EXECUTABLE=/opt/hostedtoolcache/Python/3.8.14/x64/bin/python3 ..
+make -j4 pnnx
+
+./src/pnnx || echo "pass"
+
+popd
+
+log "Test exporting to pnnx format"
+
+./conv_emformer_transducer_stateless2/export-for-ncnn.py \
+  --exp-dir $repo/exp \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  \
+  --num-encoder-layers 12 \
+  --chunk-length 32 \
+  --cnn-module-kernel 31 \
+  --left-context-length 32 \
+  --right-context-length 8 \
+  --memory-size 32
+
+./ncnn/tools/pnnx/build/src/pnnx $repo/exp/encoder_jit_trace-pnnx.pt
+./ncnn/tools/pnnx/build/src/pnnx $repo/exp/decoder_jit_trace-pnnx.pt
+./ncnn/tools/pnnx/build/src/pnnx $repo/exp/joiner_jit_trace-pnnx.pt
+
+./conv_emformer_transducer_stateless2/streaming-ncnn-decode.py \
+ --tokens $repo/data/lang_bpe_500/tokens.txt \
+ --encoder-param-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.param \
+ --encoder-bin-filename $repo/exp/encoder_jit_trace-pnnx.ncnn.bin \
+ --decoder-param-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.param \
+ --decoder-bin-filename $repo/exp/decoder_jit_trace-pnnx.ncnn.bin \
+ --joiner-param-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.param \
+ --joiner-bin-filename $repo/exp/joiner_jit_trace-pnnx.ncnn.bin \
+ $repo/test_wavs/1089-134686-0001.wav

.github/scripts/run-librispeech-lstm-transducer-stateless2-2022-09-03.sh

@@ -16,6 +16,7 @@ log "Downloading pre-trained model from $repo_url"
 git lfs install
 git clone $repo_url
 repo=$(basename $repo_url)
+abs_repo=$(realpath $repo)
 
 log "Display test files"
 tree $repo/
@@ -174,6 +175,92 @@ done
 
 echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
 echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+
+if [[ x"${GITHUB_EVENT_LABEL_NAME}" == x"shallow-fusion" ]]; then
+  lm_repo_url=https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm
+  log "Download pre-trained RNN-LM model from ${lm_repo_url}"
+  GIT_LFS_SKIP_SMUDGE=1 git clone $lm_repo_url
+  lm_repo=$(basename $lm_repo_url)
+  pushd $lm_repo
+  git lfs pull --include "exp/pretrained.pt"
+  mv exp/pretrained.pt exp/epoch-88.pt
+  popd
+
+  mkdir -p lstm_transducer_stateless2/exp
+  ln -sf $PWD/$repo/exp/pretrained.pt lstm_transducer_stateless2/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh lstm_transducer_stateless2/exp
+
+  log "Decoding test-clean and test-other with RNN LM"
+
+  ./lstm_transducer_stateless2/decode.py \
+    --use-averaged-model 0 \
+    --epoch 999 \
+    --avg 1 \
+    --exp-dir lstm_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search_lm_shallow_fusion \
+    --beam 4 \
+    --use-shallow-fusion 1 \
+    --lm-type rnn \
+    --lm-exp-dir $lm_repo/exp \
+    --lm-epoch 88 \
+    --lm-avg 1 \
+    --lm-scale 0.3 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1
+fi
+
+if [[ x"${GITHUB_EVENT_LABEL_NAME}" == x"LODR" ]]; then
+  bigram_repo_url=https://huggingface.co/marcoyang/librispeech_bigram
+  log "Download bi-gram LM from ${bigram_repo_url}"
+  GIT_LFS_SKIP_SMUDGE=1 git clone $bigram_repo_url
+  bigramlm_repo=$(basename $bigram_repo_url)
+  pushd $bigramlm_repo
+  git lfs pull --include "2gram.fst.txt"
+  cp 2gram.fst.txt $abs_repo/data/lang_bpe_500/.
+  popd
+
+  lm_repo_url=https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm
+  log "Download pre-trained RNN-LM model from ${lm_repo_url}"
+  GIT_LFS_SKIP_SMUDGE=1 git clone $lm_repo_url
+  lm_repo=$(basename $lm_repo_url)
+  pushd $lm_repo
+  git lfs pull --include "exp/pretrained.pt"
+  mv exp/pretrained.pt exp/epoch-88.pt
+  popd
+
+  mkdir -p lstm_transducer_stateless2/exp
+  ln -sf $PWD/$repo/exp/pretrained.pt lstm_transducer_stateless2/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh lstm_transducer_stateless2/exp
+
+  log "Decoding test-clean and test-other"
+
+  ./lstm_transducer_stateless2/decode.py \
+    --use-averaged-model 0 \
+    --epoch 999 \
+    --avg 1 \
+    --exp-dir lstm_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search_LODR \
+    --beam 4 \
+    --use-shallow-fusion 1 \
+    --lm-type rnn \
+    --lm-exp-dir $lm_repo/exp \
+    --lm-scale 0.4 \
+    --lm-epoch 88 \
+    --rnn-lm-avg 1 \
+    --rnn-lm-num-layers 3 \
+    --rnn-lm-tie-weights 1 \
+    --tokens-ngram 2 \
+    --ngram-lm-scale -0.16
+fi
+
 if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
   mkdir -p lstm_transducer_stateless2/exp
   ln -s $PWD/$repo/exp/pretrained.pt lstm_transducer_stateless2/exp/epoch-999.pt

.github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh

@@ -83,4 +83,5 @@ if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" ==
   done
 
   rm pruned_transducer_stateless2/exp/*.pt
+  rm -r data/lang_bpe_500
 fi

.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh

@@ -82,4 +82,5 @@ if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" ==
   done
 
   rm pruned_transducer_stateless3/exp/*.pt
+  rm -r data/lang_bpe_500
 fi

.github/scripts/run-librispeech-pruned-transducer-stateless7-2022-11-11.sh

@@ -0,0 +1,137 @@
+#!/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://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
+
+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/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Test exporting to ONNX format"
+./pruned_transducer_stateless7/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --onnx 1
+
+log "Export to torchscript model"
+./pruned_transducer_stateless7/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with ONNX models"
+
+./pruned_transducer_stateless7/onnx_check.py \
+  --jit-filename $repo/exp/cpu_jit.pt \
+  --onnx-encoder-filename $repo/exp/encoder.onnx \
+  --onnx-decoder-filename $repo/exp/decoder.onnx \
+  --onnx-joiner-filename $repo/exp/joiner.onnx \
+  --onnx-joiner-encoder-proj-filename $repo/exp/joiner_encoder_proj.onnx \
+  --onnx-joiner-decoder-proj-filename $repo/exp/joiner_decoder_proj.onnx
+
+./pruned_transducer_stateless7/onnx_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --encoder-model-filename $repo/exp/encoder.onnx \
+  --decoder-model-filename $repo/exp/decoder.onnx \
+  --joiner-model-filename $repo/exp/joiner.onnx \
+  --joiner-encoder-proj-model-filename $repo/exp/joiner_encoder_proj.onnx \
+  --joiner-decoder-proj-model-filename $repo/exp/joiner_decoder_proj.onnx \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless7/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless7/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless7/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    $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 pruned_transducer_stateless7/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless7/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless7/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"
+
+    ./pruned_transducer_stateless7/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless7/exp
+  done
+
+  rm pruned_transducer_stateless7/exp/*.pt
+fi

.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-2022-12-01.sh

@@ -0,0 +1,151 @@
+#!/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://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01
+
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+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/LG.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/lm/G_4_gram.pt"
+git lfs pull --include "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./pruned_transducer_stateless7_ctc/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless7_ctc/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for m in ctc-decoding 1best; do
+  ./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
+    --model-filename $repo/exp/cpu_jit.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --G $repo/data/lm/G_4_gram.pt \
+    --method $m \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless7_ctc/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless7_ctc/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for m in ctc-decoding 1best; do
+  ./pruned_transducer_stateless7_ctc/pretrained_ctc.py \
+    --checkpoint $repo/exp/pretrained.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --G $repo/data/lm/G_4_gram.pt \
+    --method $m \
+    --sample-rate 16000 \
+    $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 pruned_transducer_stateless7_ctc/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless7_ctc/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless7_ctc/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"
+
+    ./pruned_transducer_stateless7_ctc/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless7_ctc/exp
+  done
+
+  for m in ctc-decoding 1best; do
+    ./pruned_transducer_stateless7_ctc/ctc_decode.py \
+        --epoch 999 \
+        --avg 1 \
+        --exp-dir ./pruned_transducer_stateless7_ctc/exp \
+        --max-duration $max_duration \
+        --use-averaged-model 0 \
+        --decoding-method $m \
+        --hlg-scale 0.6 \
+        --lm-dir data/lm
+  done
+
+  rm pruned_transducer_stateless7_ctc/exp/*.pt
+fi

.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2022-12-15.sh

@@ -0,0 +1,148 @@
+#!/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://huggingface.co/yfyeung/icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2022-12-14
+
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+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/LG.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 "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./pruned_transducer_stateless7_ctc_bs/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless7_ctc_bs/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for m in ctc-decoding 1best; do
+  ./pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py \
+    --model-filename $repo/exp/cpu_jit.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --method $m \
+    --sample-rate 16000 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless7_ctc_bs/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless7_ctc_bs/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for m in ctc-decoding 1best; do
+  ./pruned_transducer_stateless7_ctc_bs/pretrained_ctc.py \
+    --checkpoint $repo/exp/pretrained.pt \
+    --words-file $repo/data/lang_bpe_500/words.txt  \
+    --HLG $repo/data/lang_bpe_500/HLG.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --method $m \
+    --sample-rate 16000 \
+    $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 pruned_transducer_stateless7_ctc_bs/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless7_ctc_bs/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless7_ctc_bs/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"
+
+    ./pruned_transducer_stateless7_ctc_bs/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless7_ctc_bs/exp
+  done
+
+  for m in ctc-decoding 1best; do
+    ./pruned_transducer_stateless7_ctc_bs/ctc_decode.py \
+        --epoch 999 \
+        --avg 1 \
+        --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+        --max-duration $max_duration \
+        --use-averaged-model 0 \
+        --decoding-method $m \
+        --hlg-scale 0.6
+  done
+  
+  rm pruned_transducer_stateless7_ctc_bs/exp/*.pt
+fi

.github/scripts/run-librispeech-pruned-transducer-stateless7-streaming-2022-12-29.sh

@@ -0,0 +1,148 @@
+#!/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://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+
+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/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+git lfs pull --include "exp/encoder_jit_trace.pt"
+git lfs pull --include "exp/decoder_jit_trace.pt"
+git lfs pull --include "exp/joiner_jit_trace.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Export to torchscript model"
+./pruned_transducer_stateless7_streaming/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --decode-chunk-len 32 \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless7_streaming/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  --decode-chunk-len 32 \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "Export to torchscript model by torch.jit.trace()"
+./pruned_transducer_stateless7_streaming/jit_trace_export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --decode-chunk-len 32 \
+  --epoch 99 \
+  --avg 1
+
+log "Decode with models exported by torch.jit.trace()"
+
+./pruned_transducer_stateless7_streaming/jit_trace_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --encoder-model-filename $repo/exp/encoder_jit_trace.pt \
+  --decoder-model-filename $repo/exp/decoder_jit_trace.pt \
+  --joiner-model-filename $repo/exp/joiner_jit_trace.pt \
+  --decode-chunk-len 32 \
+  $repo/test_wavs/1089-134686-0001.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless7_streaming/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --decode-chunk-len 32 \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless7_streaming/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --decode-chunk-len 32 \
+    $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 pruned_transducer_stateless7_streaming/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless7_streaming/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless7_streaming/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+  num_decode_stream=200
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "decoding with $method"
+
+    ./pruned_transducer_stateless7_streaming/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --decode-chunk-len 32 \
+      --exp-dir pruned_transducer_stateless7_streaming/exp
+  done
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless7_streaming/streaming_decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --decode-chunk-len 32 \
+      --num-decode-streams $num_decode_stream
+      --exp-dir pruned_transducer_stateless7_streaming/exp
+  done
+
+  rm pruned_transducer_stateless7_streaming/exp/*.pt
+fi

.github/scripts/run-librispeech-pruned-transducer-stateless8-2022-11-14.sh

@@ -0,0 +1,116 @@
+#!/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://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless8-2022-11-14
+
+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/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/cpu_jit.pt"
+git lfs pull --include "exp/pretrained.pt"
+ln -s pretrained.pt epoch-99.pt
+ls -lh *.pt
+popd
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless8/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+log "Export to torchscript model"
+./pruned_transducer_stateless8/export.py \
+  --exp-dir $repo/exp \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --use-averaged-model false \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./pruned_transducer_stateless8/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless8/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    $repo/test_wavs/1089-134686-0001.wav \
+    $repo/test_wavs/1221-135766-0001.wav \
+    $repo/test_wavs/1221-135766-0002.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless8/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    $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 pruned_transducer_stateless8/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless8/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless8/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"
+
+    ./pruned_transducer_stateless8/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless8/exp
+  done
+
+  rm pruned_transducer_stateless8/exp/*.pt
+fi

.github/scripts/run-librispeech-zipformer-mmi-2022-12-08.sh

@@ -0,0 +1,103 @@
+#!/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://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-mmi-2022-12-08
+
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+git lfs pull --include "data/lang_bpe_500/3gram.pt"
+git lfs pull --include "data/lang_bpe_500/4gram.pt"
+git lfs pull --include "data/lang_bpe_500/L.pt"
+git lfs pull --include "data/lang_bpe_500/LG.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 "exp/cpu_jit.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_mmi/export.py \
+  --exp-dir $repo/exp \
+  --use-averaged-model false \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --jit 1
+
+ls -lh $repo/exp/*.pt
+
+log "Decode with models exported by torch.jit.script()"
+
+./zipformer_mmi/jit_pretrained.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --nn-model-filename $repo/exp/cpu_jit.pt \
+  --lang-dir $repo/data/lang_bpe_500 \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+for method in 1best nbest nbest-rescoring-LG nbest-rescoring-3-gram nbest-rescoring-4-gram; do
+  log "$method"
+
+  ./zipformer_mmi/pretrained.py \
+    --method $method \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_bpe_500 \
+    --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    $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_mmi/exp
+  ln -s $PWD/$repo/exp/pretrained.pt zipformer_mmi/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_bpe_500 data/
+
+  ls -lh data
+  ls -lh zipformer_mmi/exp
+
+  log "Decoding test-clean and test-other"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in 1best nbest nbest-rescoring-LG nbest-rescoring-3-gram nbest-rescoring-4-gram; do
+    log "Decoding with $method"
+
+    ./zipformer_mmi/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --use-averaged-model 0 \
+      --nbest-scale 1.2 \
+      --hp-scale 1.0 \
+      --max-duration $max_duration \
+      --lang-dir $repo/data/lang_bpe_500 \
+      --exp-dir zipformer_mmi/exp
+  done
+
+  rm zipformer_mmi/exp/*.pt
+fi

.github/workflows/build-doc.yml

@@ -26,6 +26,10 @@ on:
   pull_request:
     types: [labeled]
 
+concurrency:
+  group: build_doc-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   build-doc:
     if: github.event.label.name == 'doc' || github.event_name == 'push'

.github/workflows/run-aishell-2022-06-20.yml

@@ -34,6 +34,10 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+concurrency:
+  group: run_aishell_2022_06_20-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_aishell_2022_06_20:
     if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'

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

@@ -33,6 +33,10 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+concurrency:
+  group: run_gigaspeech_2022_05_13-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_gigaspeech_2022_05_13:
     if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'

.github/workflows/run-librispeech-2022-03-12.yml

@@ -33,6 +33,10 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+concurrency:
+  group: run_librispeech_2022_03_12-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_librispeech_2022_03_12:
     if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'

.github/workflows/run-librispeech-2022-04-29.yml

@@ -33,6 +33,10 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+concurrency:
+  group: run_librispeech_2022_04_29-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_librispeech_2022_04_29:
     if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'

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

@@ -33,6 +33,10 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+concurrency:
+  group: run_librispeech_2022_05_13-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_librispeech_2022_05_13:
     if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'

.github/workflows/run-librispeech-2022-11-11-stateless7.yml

@@ -0,0 +1,159 @@
+# 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-librispeech-2022-11-11-stateless7
+# 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_librispeech_2022_11_11_zipformer-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_11_11_zipformer:
+    if: github.event.label.name == 'onnx' || 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: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2022-09-25
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree sox
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless7-2022-11-11.sh
+
+      - name: Display decoding results for librispeech pruned_transducer_stateless7
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless7/exp
+
+          cd pruned_transducer_stateless7
+          echo "results for pruned_transducer_stateless7"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech pruned_transducer_stateless7
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless7-2022-11-11
+          path: egs/librispeech/ASR/pruned_transducer_stateless7/exp/

.github/workflows/run-librispeech-2022-11-14-stateless8.yml

@@ -0,0 +1,159 @@
+# 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-librispeech-2022-11-14-stateless8
+# 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_librispeech_2022_11_14_zipformer_stateless8-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_11_14_zipformer_stateless8:
+    if: 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: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2022-09-25
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree sox
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless8-2022-11-14.sh
+
+      - name: Display decoding results for librispeech pruned_transducer_stateless8
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless8/exp
+
+          cd pruned_transducer_stateless8
+          echo "results for pruned_transducer_stateless8"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech pruned_transducer_stateless8
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless8-2022-11-14
+          path: egs/librispeech/ASR/pruned_transducer_stateless8/exp/

.github/workflows/run-librispeech-2022-12-01-stateless7-ctc.yml

@@ -0,0 +1,163 @@
+# 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-librispeech-2022-12-01-stateless7-ctc
+# 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 * * *"
+
+jobs:
+  run_librispeech_2022_11_11_zipformer:
+    if: 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: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2022-09-25
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree sox
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-2022-12-01.sh
+
+      - name: Display decoding results for librispeech pruned_transducer_stateless7_ctc
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless7_ctc/exp
+
+          cd pruned_transducer_stateless7_ctc
+          echo "results for pruned_transducer_stateless7_ctc"
+          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
+
+          echo "===ctc decoding==="
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===1best==="
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech pruned_transducer_stateless7_ctc
+        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-18.04-cpu-pruned_transducer_stateless7-ctc-2022-12-01
+          path: egs/librispeech/ASR/pruned_transducer_stateless7_ctc/exp/

.github/workflows/run-librispeech-2022-12-08-zipformer-mmi.yml

@@ -0,0 +1,167 @@
+# Copyright      2022  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.
+
+name: run-librispeech-2022-12-08-zipformer-mmi
+# 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_librispeech_2022_12_08_zipformer-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_12_08_zipformer:
+    if: 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: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2022-09-25
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree sox
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-zipformer-mmi-2022-12-08.sh
+
+      - name: Display decoding results for librispeech zipformer-mmi
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./zipformer-mmi/exp
+
+          cd zipformer-mmi
+          echo "results for zipformer-mmi"
+          echo "===1best==="
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===nbest==="
+          find exp/nbest -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/nbest -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===nbest-rescoring-LG==="
+          find exp/nbest-rescoring-LG -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/nbest-rescoring-LG -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===nbest-rescoring-3-gram==="
+          find exp/nbest-rescoring-3-gram -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/nbest-rescoring-3-gram -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===nbest-rescoring-4-gram==="
+          find exp/nbest-rescoring-4-gram -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/nbest-rescoring-4-gram -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech zipformer-mmi
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-zipformer_mmi-2022-12-08
+          path: egs/librispeech/ASR/zipformer_mmi/exp/

.github/workflows/run-librispeech-2022-12-15-stateless7-ctc-bs.yml

@@ -0,0 +1,163 @@
+# 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-librispeech-2022-12-15-stateless7-ctc-bs
+# 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 * * *"
+
+jobs:
+  run_librispeech_2022_12_15_zipformer_ctc_bs:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event.label.name == 'blank-skip' || 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: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2022-09-25
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree sox
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-bs-2022-12-15.sh
+
+      - name: Display decoding results for librispeech pruned_transducer_stateless7_ctc_bs
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless7_ctc_bs/exp
+
+          cd pruned_transducer_stateless7_ctc_bs
+          echo "results for pruned_transducer_stateless7_ctc_bs"
+          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
+
+          echo "===ctc decoding==="
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===1best==="
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech pruned_transducer_stateless7_ctc_bs
+        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-18.04-cpu-pruned_transducer_stateless7-ctc-bs-2022-12-15
+          path: egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/exp/

.github/workflows/run-librispeech-2022-12-29-stateless7-streaming.yml

@@ -0,0 +1,172 @@
+# 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-librispeech-2022-12-29-stateless7-streaming
+# 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_librispeech_2022_12_29_zipformer_streaming-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_12_29_zipformer_streaming:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event.label.name == 'streaming-zipformer' || 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: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2022-09-25
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree sox
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-pruned-transducer-stateless7-streaming-2022-12-29.sh
+
+      - name: Display decoding results for librispeech pruned_transducer_stateless7_streaming
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./pruned_transducer_stateless7_streaming/exp
+
+          cd pruned_transducer_stateless7_streaming
+          echo "results for pruned_transducer_stateless7_streaming"
+          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
+
+          echo "===streaming greedy search==="
+          find exp/streaming/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/streaming/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===streaming fast_beam_search==="
+          find exp/streaming/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/streaming/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===streaming modified beam search==="
+          find exp/streaming/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/streaming/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+
+      - name: Upload decoding results for librispeech pruned_transducer_stateless7_streaming
+        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-18.04-cpu-pruned_transducer_stateless7-streaming-2022-12-29
+          path: egs/librispeech/ASR/pruned_transducer_stateless7_streaming/exp/

.github/workflows/run-librispeech-conformer-ctc3-2022-11-28.yml

@@ -0,0 +1,155 @@
+# 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-librispeech-conformer-ctc3-2022-11-28
+# 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_librispeech_2022_11_28_conformer_ctc3-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_librispeech_2022_11_28_conformer_ctc3:
+    if: 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: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2022-09-25
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Cache LibriSpeech test-clean and test-other datasets
+        id: libri-test-clean-and-test-other-data
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/download
+          key: cache-libri-test-clean-and-test-other
+
+      - name: Download LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
+
+      - name: Prepare manifests for LibriSpeech test-clean and test-other
+        shell: bash
+        run: |
+          .github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
+
+      - name: Cache LibriSpeech test-clean and test-other fbank features
+        id: libri-test-clean-and-test-other-fbank
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/fbank-libri
+          key: cache-libri-fbank-test-clean-and-test-other-v2
+
+      - name: Compute fbank for LibriSpeech test-clean and test-other
+        if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree sox
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-conformer-ctc3-2022-11-28.sh
+
+      - name: Display decoding results for librispeech conformer_ctc3
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR/
+          tree ./conformer_ctc3/exp
+
+          cd conformer_ctc3
+          echo "results for conformer_ctc3"
+          echo "===ctc-decoding==="
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+          echo "===1best==="
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find exp/1best -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Upload decoding results for librispeech conformer_ctc3
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-conformer_ctc3-2022-11-28
+          path: egs/librispeech/ASR/conformer_ctc3/exp/

.github/workflows/run-librispeech-conv-emformer-transducer-stateless2-2022-12-05.yml

@@ -0,0 +1,77 @@
+name: run-librispeech-conv-emformer-transducer-stateless2-2022-12-05
+
+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 * * *"
+
+jobs:
+  run_librispeech_conv_emformer_transducer_stateless2_2022_12_05:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'ncnn' || 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: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | grep -v kaldifst | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}-2022-09-25
+
+      - 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: ${{ github.event.label.name }}
+        run: |
+          mkdir -p egs/librispeech/ASR/data
+          ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
+          ls -lh egs/librispeech/ASR/data/*
+
+          sudo apt-get -qq install git-lfs tree sox
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-librispeech-conv-emformer-transducer-stateless2-2022-12-05.sh

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

@@ -16,9 +16,13 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+concurrency:
+  group: run_librispeech_lstm_transducer_stateless2_2022_09_03-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_librispeech_lstm_transducer_stateless2_2022_09_03:
-    if: github.event.label.name == 'ready' || github.event.label.name == 'ncnn' || github.event.label.name == 'onnx' || github.event_name == 'push' || github.event_name == 'schedule'
+    if: github.event.label.name == 'ready' || github.event.label.name == 'LODR' || github.event.label.name == 'shallow-fusion' || github.event.label.name == 'ncnn' || github.event.label.name == 'onnx' || github.event_name == 'push' || github.event_name == 'schedule'
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
@@ -107,7 +111,7 @@ jobs:
           export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
           export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
 
-          .github/scripts/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml
+          .github/scripts/run-librispeech-lstm-transducer-stateless2-2022-09-03.sh
 
       - name: Display decoding results for lstm_transducer_stateless2
         if: github.event_name == 'schedule'
@@ -128,9 +132,32 @@ jobs:
           find modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
           find modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
 
+      - name: Display decoding results for lstm_transducer_stateless2
+        if: github.event.label.name == 'shallow-fusion'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR
+          tree lstm_transducer_stateless2/exp
+          cd lstm_transducer_stateless2/exp
+          echo "===modified_beam_search_lm_shallow_fusion==="
+          echo "===Using RNNLM==="
+          find modified_beam_search_lm_shallow_fusion  -name "log-*rnn*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          find modified_beam_search_lm_shallow_fusion  -name "log-*rnn*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
+
+      - name: Display decoding results for lstm_transducer_stateless2
+        if: github.event.label.name == 'LODR'
+        shell: bash
+        run: |
+          cd egs/librispeech/ASR
+          tree lstm_transducer_stateless2/exp
+          cd lstm_transducer_stateless2/exp
+          echo "===modified_beam_search_rnnlm_LODR==="
+          find modified_beam_search_LODR  -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
+          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
-        if: github.event_name == 'schedule'
+        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-18.04-cpu-lstm_transducer_stateless2-2022-09-03
           path: egs/librispeech/ASR/lstm_transducer_stateless2/exp/

.github/workflows/run-librispeech-pruned-transducer-stateless3-2022-05-13.yml

@@ -33,6 +33,10 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+concurrency:
+  group: run_librispeech_pruned_transducer_stateless3_2022_05_13-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_librispeech_pruned_transducer_stateless3_2022_05_13:
     if: github.event.label.name == 'onnx' || github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'

.github/workflows/run-librispeech-streaming-transducer-stateless2-2022-06-26.yml

@@ -33,6 +33,10 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+concurrency:
+  group: run_librispeech_streaming_2022_06_26-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_librispeech_streaming_2022_06_26:
     if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'

.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml

@@ -33,6 +33,10 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+concurrency:
+  group: run_librispeech_2022_04_19-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_librispeech_2022_04_19:
     if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'

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

@@ -23,6 +23,10 @@ on:
   pull_request:
     types: [labeled]
 
+concurrency:
+  group: run_pre_trained_conformer_ctc-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_pre_trained_conformer_ctc:
     if: github.event.label.name == 'ready' || github.event_name == 'push'

.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml

@@ -32,6 +32,10 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+concurrency:
+  group: run_pre_trained_transducer_stateless_multi_datasets_librispeech_100h-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_pre_trained_transducer_stateless_multi_datasets_librispeech_100h:
     if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'

.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml

@@ -32,6 +32,10 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+concurrency:
+  group: run_pre_trained_transducer_stateless_multi_datasets_librispeech_960h-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_pre_trained_transducer_stateless_multi_datasets_librispeech_960h:
     if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'

.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml

@@ -23,6 +23,10 @@ on:
   pull_request:
     types: [labeled]
 
+concurrency:
+  group: run_pre_trained_transducer_stateless_modified_2_aishell-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_pre_trained_transducer_stateless_modified_2_aishell:
     if: github.event.label.name == 'ready' || github.event_name == 'push'

.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml

@@ -23,6 +23,10 @@ on:
   pull_request:
     types: [labeled]
 
+concurrency:
+  group: run_pre_trained_transducer_stateless_modified_aishell-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_pre_trained_transducer_stateless_modified_aishell:
     if: github.event.label.name == 'ready' || github.event_name == 'push'

.github/workflows/run-pretrained-transducer-stateless.yml

@@ -32,6 +32,10 @@ on:
     # nightly build at 15:50 UTC time every day
     - cron: "50 15 * * *"
 
+concurrency:
+  group: run_pre_trained_transducer_stateless-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_pre_trained_transducer_stateless:
     if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'

.github/workflows/run-pretrained-transducer.yml

@@ -23,6 +23,10 @@ on:
   pull_request:
     types: [labeled]
 
+concurrency:
+  group: run_pre_trained_transducer-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run_pre_trained_transducer:
     if: github.event.label.name == 'ready' || github.event_name == 'push'

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

@@ -0,0 +1,71 @@
+name: run-ptb-rnn-lm-training
+
+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_ptb_rnn_lm_training-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  run_ptb_rnn_lm_training:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'rnnlm' || 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: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | grep -v kaldifst | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Prepare data
+        shell: bash
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          cd egs/ptb/LM
+          ./prepare.sh
+
+      - name: Run training
+        shell: bash
+        run: |
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          cd egs/ptb/LM
+          ./train-rnn-lm.sh --world-size 1 --num-epochs 5 --use-epoch 4 --use-avg 2
+
+      - name: Upload pretrained models
+        uses: actions/upload-artifact@v2
+        if: github.event.label.name == 'ready' || github.event.label.name == 'rnnlm' || github.event_name == 'push' || github.event_name == 'schedule'
+        with:
+          name: python-${{ matrix.python-version }}-ubuntu-rnn-lm-ptb
+          path: egs/ptb/LM/my-rnnlm-exp/

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

@@ -23,8 +23,12 @@ on:
   pull_request:
     types: [labeled]
 
+concurrency:
+  group: run_wenetspeech_pruned_transducer_stateless2-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
-  run_librispeech_pruned_transducer_stateless3_2022_05_13:
+  run_wenetspeech_pruned_transducer_stateless2:
     if: github.event.label.name == 'onnx' || github.event.label.name == 'ready' || github.event_name == 'push' || github.event.label.name == 'wenetspeech'
     runs-on: ${{ matrix.os }}
     strategy:

.github/workflows/run-yesno-recipe.yml

@@ -21,11 +21,15 @@ on:
     branches:
       - master
   pull_request:
-    types: [labeled]
+    branches:
+      - master
+
+concurrency:
+  group: run-yesno-recipe-${{ github.ref }}
+  cancel-in-progress: true
 
 jobs:
   run-yesno-recipe:
-    if: github.event.label.name == 'ready' || github.event_name == 'push'
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
@@ -61,7 +65,7 @@ jobs:
 
       - name: Install Python dependencies
         run: |
-          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          grep -v '^#' ./requirements-ci.txt  | grep -v kaldifst | xargs -n 1 -L 1 pip install
           pip uninstall -y protobuf
           pip install --no-binary protobuf protobuf
 

.github/workflows/style_check.yml

@@ -24,6 +24,10 @@ on:
     branches:
       - master
 
+concurrency:
+  group: style_check-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   style_check:
     runs-on: ${{ matrix.os }}
@@ -45,17 +49,18 @@ jobs:
 
       - name: Install Python dependencies
         run: |
-          python3 -m pip install --upgrade pip black==21.6b0 flake8==3.9.2 click==8.0.4
-          # See https://github.com/psf/black/issues/2964
-          # The version of click should be selected from 8.0.0, 8.0.1, 8.0.2, 8.0.3, and 8.0.4
+          python3 -m pip install --upgrade pip black==22.3.0 flake8==5.0.4 click==8.1.0
+          # Click issue fixed in https://github.com/psf/black/pull/2966
 
       - name: Run flake8
         shell: bash
         working-directory: ${{github.workspace}}
         run: |
           # stop the build if there are Python syntax errors or undefined names
-          flake8 . --count --show-source --statistics
-          flake8 .
+          flake8 . --count --select=E9,F63,F7,F82 --show-source --statistics
+          # exit-zero treats all errors as warnings. The GitHub editor is 127 chars wide
+          flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 \
+            --statistics --extend-ignore=E203,E266,E501,F401,E402,F403,F841,W503
 
       - name: Run black
         shell: bash

.github/workflows/test.yml

@@ -21,26 +21,23 @@ on:
     branches:
       - master
   pull_request:
-    types: [labeled]
+    branches:
+      - master
+
+concurrency:
+  group: test-${{ github.ref }}
+  cancel-in-progress: true
 
 jobs:
   test:
-    if: github.event.label.name == 'ready' || github.event_name == 'push'
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
-        # os: [ubuntu-18.04, macos-10.15]
-        # disable macOS test for now.
-        os: [ubuntu-18.04]
-        python-version: [3.7, 3.8]
-        torch: ["1.8.0", "1.11.0"]
-        torchaudio: ["0.8.0", "0.11.0"]
-        k2-version: ["1.15.1.dev20220427"]
-        exclude:
-          - torch: "1.8.0"
-            torchaudio: "0.11.0"
-          - torch: "1.11.0"
-            torchaudio: "0.8.0"
+        os: [ubuntu-latest]
+        python-version: ["3.8"]
+        torch: ["1.10.0"]
+        torchaudio: ["0.10.0"]
+        k2-version: ["1.23.2.dev20221201"]
 
       fail-fast: false
 
@@ -67,11 +64,7 @@ jobs:
           # numpy 1.20.x does not support python 3.6
           pip install numpy==1.19
           pip install torch==${{ matrix.torch }}+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
-          if [[ ${{ matrix.torchaudio }} == "0.11.0" ]]; then
           pip install torchaudio==${{ matrix.torchaudio }}+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
-          else
-            pip install torchaudio==${{ matrix.torchaudio }}
-          fi
 
           pip install k2==${{ matrix.k2-version }}+cpu.torch${{ matrix.torch }} -f https://k2-fsa.org/nightly/
           pip install git+https://github.com/lhotse-speech/lhotse
@@ -79,6 +72,8 @@ jobs:
           pip uninstall -y protobuf
           pip install --no-binary protobuf protobuf
 
+          pip install kaldifst
+          pip install onnxruntime
           pip install -r requirements.txt
 
       - name: Install graphviz
@@ -118,10 +113,12 @@ jobs:
           cd ../pruned_transducer_stateless4
           pytest -v -s
 
+          cd ../pruned_transducer_stateless7
+          pytest -v -s
+
           cd ../transducer_stateless
           pytest -v -s
 
-          if [[ ${{ matrix.torchaudio }} == "0.10.0" ]]; then
           cd ../transducer
           pytest -v -s
 
@@ -130,7 +127,6 @@ jobs:
 
           cd ../transducer_lstm
           pytest -v -s
-          fi
 
       - name: Run tests
         if: startsWith(matrix.os, 'macos')
@@ -161,7 +157,6 @@ jobs:
           cd ../transducer_stateless
           pytest -v -s
 
-          if [[ ${{ matrix.torchaudio }} == "0.10.0" ]]; then
           cd ../transducer
           pytest -v -s
 
@@ -170,4 +165,3 @@ jobs:
 
           cd ../transducer_lstm
           pytest -v -s
-          fi

.gitignore

@@ -11,5 +11,26 @@ log
 *.bak
 *-bak
 *bak.py
+
+# Ignore Mac system files
+.DS_store
+
+# Ignore node_modules folder
+node_modules
+
+# ignore .nfs
+
+.nfs*
+
+# Ignore all text files
+*.txt
+
+# Ignore files related to API keys
+.env
+
+# Ignore SASS config files
+.sass-cache
+
 *.param
 *.bin
+.DS_Store

.pre-commit-config.yaml

@@ -1,26 +1,38 @@
 repos:
   - repo: https://github.com/psf/black
-    rev: 21.6b0
+    rev: 22.3.0
     hooks:
       - id: black
-        args: [--line-length=80]
-        additional_dependencies: ['click==8.0.1']
+        args: ["--line-length=88"]
+        additional_dependencies: ['click==8.1.0']
         exclude: icefall\/__init__\.py
 
   - repo: https://github.com/PyCQA/flake8
-    rev: 3.9.2
+    rev: 5.0.4
     hooks:
       - id: flake8
-        args: [--max-line-length=80]
+        args: ["--max-line-length=88", "--extend-ignore=E203,E266,E501,F401,E402,F403,F841,W503"]
+
+      # What are we ignoring here?
+      # E203: whitespace before ':'
+      # E266: too many leading '#' for block comment
+      # E501: line too long
+      # F401: module imported but unused
+      # E402: module level import not at top of file
+      # F403: 'from module import *' used; unable to detect undefined names
+      # F841: local variable is assigned to but never used
+      # W503: line break before binary operator
+      # In addition, the default ignore list is:
+      # E121,E123,E126,E226,E24,E704,W503,W504
 
   - repo: https://github.com/pycqa/isort
-    rev: 5.9.2
+    rev: 5.10.1
     hooks:
       - id: isort
-        args: [--profile=black, --line-length=80]
+        args: ["--profile=black"]
 
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.0.1
+    rev: v4.2.0
     hooks:
       - id: check-executables-have-shebangs
       - id: end-of-file-fixer

README.md

@@ -82,7 +82,7 @@ The WER for this model is:
 |-----|------------|------------|
 | WER | 6.59       | 17.69      |
 
-We provide a Colab notebook to run a pre-trained TDNN LSTM CTC model:  [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1kNmDXNMwREi0rZGAOIAOJo93REBuOTcd?usp=sharing)
+We provide a Colab notebook to run a pre-trained TDNN LSTM CTC model:  [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1-iSfQMp2So-We_Uu49N4AAcMInB72u9z?usp=sharing)
 
 
 #### Transducer: Conformer encoder + LSTM decoder
@@ -162,7 +162,7 @@ The CER for this model is:
 |-----|-------|
 | CER | 10.16 |
 
-We provide a Colab notebook to run a pre-trained TDNN LSTM CTC model:  [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1qULaGvXq7PCu_P61oubfz9b53JzY4H3z?usp=sharing)
+We provide a Colab notebook to run a pre-trained TDNN LSTM CTC model:  [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1jbyzYq3ytm6j2nlEt-diQm-6QVWyDDEa?usp=sharing)
 
 ### TIMIT
 

docker/README.md

@@ -72,14 +72,14 @@ docker run -it --runtime=nvidia --shm-size=2gb --name=icefall --gpus all icefall
 ```
 
 ### Tips:
-1. Since your data and models most probably won't be in the docker, you must use the -v flag to access the host machine. Do this by specifying `-v {/path/in/docker}:{/path/in/host/machine}`. 
+1. Since your data and models most probably won't be in the docker, you must use the -v flag to access the host machine. Do this by specifying `-v {/path/in/host/machine}:{/path/in/docker}`.
 
 2. Also, if your environment requires a proxy, this would be a good time to add it in too: `-e http_proxy=http://aaa.bb.cc.net:8080 -e https_proxy=http://aaa.bb.cc.net:8080`.
 
 Overall, your docker run command should look like this.
 
 ```bash
-docker run -it --runtime=nvidia --shm-size=2gb --name=icefall --gpus all -v {/path/in/docker}:{/path/in/host/machine} -e http_proxy=http://aaa.bb.cc.net:8080 -e https_proxy=http://aaa.bb.cc.net:8080 icefall/pytorch1.12.1
+docker run -it --runtime=nvidia --shm-size=2gb --name=icefall --gpus all -v {/path/in/host/machine}:{/path/in/docker} -e http_proxy=http://aaa.bb.cc.net:8080 -e https_proxy=http://aaa.bb.cc.net:8080 icefall/pytorch1.12.1
 ```
 
 You can explore more docker run options [here](https://docs.docker.com/engine/reference/commandline/run/) to suit your environment.

docker/Ubuntu18.04-pytorch1.12.1-cuda11.3-cudnn8/Dockerfile

@@ -51,8 +51,9 @@ RUN wget -P /opt https://downloads.xiph.org/releases/flac/flac-1.3.2.tar.xz  &&
     find /opt/flac-1.3.2  -type f \( -name "*.o" -o -name "*.la" -o -name "*.a" \) -exec rm {} \; && \
     cd -
 
-RUN pip install kaldiio graphviz && \
-	conda install -y -c pytorch torchaudio
+RUN conda install -y -c pytorch torchaudio=0.12 && \
+    pip install graphviz
+
 
 #install k2 from source
 RUN git clone https://github.com/k2-fsa/k2.git /opt/k2 && \
@@ -67,6 +68,7 @@ RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
 	cd /workspace/icefall && \
 	pip install -r requirements.txt
 
+RUN pip install kaldifeat
 ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
 
 WORKDIR /workspace/icefall

docker/Ubuntu18.04-pytorch1.7.1-cuda11.0-cudnn8/Dockerfile

@@ -69,8 +69,8 @@ RUN wget -P /opt https://downloads.xiph.org/releases/flac/flac-1.3.2.tar.xz  &&
     find /opt/flac-1.3.2  -type f \( -name "*.o" -o -name "*.la" -o -name "*.a" \) -exec rm {} \; && \
     cd -
 
-RUN pip install kaldiio graphviz && \
-	conda install -y -c pytorch torchaudio=0.7.1
+RUN conda install -y -c pytorch torchaudio=0.7.1 && \
+    pip install graphviz
 
 #install k2 from source
 RUN git clone https://github.com/k2-fsa/k2.git /opt/k2 && \
@@ -88,4 +88,3 @@ RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \
 ENV PYTHONPATH /workspace/icefall:$PYTHONPATH
 
 WORKDIR /workspace/icefall
-

docs/README.md

@@ -0,0 +1,24 @@
+
+## Usage
+
+```bash
+cd /path/to/icefall/docs
+pip install -r requirements.txt
+make clean
+make html
+cd build/html
+python3 -m http.server 8000
+```
+
+It prints:
+
+```
+Serving HTTP on 0.0.0.0 port 8000 (http://0.0.0.0:8000/) ...
+```
+
+Open your browser and go to <http://0.0.0.0:8000/> to view the generated
+documentation.
+
+Done!
+
+**Hint**: You can change the port number when starting the server.

docs/source/conf.py

@@ -78,3 +78,12 @@ html_context = {
 }
 
 todo_include_todos = True
+
+rst_epilog = """
+.. _sherpa-ncnn: https://github.com/k2-fsa/sherpa-ncnn
+.. _icefall: https://github.com/k2-fsa/icefall
+.. _git-lfs: https://git-lfs.com/
+.. _ncnn: https://github.com/tencent/ncnn
+.. _LibriSpeech: https://www.openslr.org/12
+.. _musan: http://www.openslr.org/17/
+"""

docs/source/contributing/code-style.rst

@@ -11,9 +11,9 @@ We use the following tools to make the code style to be as consistent as possibl
 
 The following versions of the above tools are used:
 
-  - ``black == 12.6b0``
-  - ``flake8 == 3.9.2``
-  - ``isort == 5.9.2``
+  - ``black == 22.3.0``
+  - ``flake8 == 5.0.4``
+  - ``isort == 5.10.1``
 
 After running the following commands:
 
@@ -54,10 +54,17 @@ it should succeed this time:
 If you want to check the style of your code before ``git commit``, you
 can do the following:
 
+  .. code-block:: bash
+
+    $ pre-commit install
+    $ pre-commit run
+
+Or without installing the pre-commit hooks:
+
   .. code-block:: bash
 
     $ cd icefall
-    $ pip install black==21.6b0 flake8==3.9.2 isort==5.9.2
+    $ pip install black==22.3.0 flake8==5.0.4 isort==5.10.1
     $ black --check your_changed_file.py
     $ black your_changed_file.py  # modify it in-place
     $

docs/source/faqs.rst

@@ -0,0 +1,107 @@
+Frequently Asked Questions (FAQs)
+=================================
+
+In this section, we collect issues reported by users and post the corresponding
+solutions.
+
+
+OSError: libtorch_hip.so: cannot open shared object file: no such file or directory
+-----------------------------------------------------------------------------------
+
+One user is using the following code to install ``torch`` and ``torchaudio``:
+
+.. code-block:: bash
+
+  pip install \
+    torch==1.10.0+cu111 \
+    torchvision==0.11.0+cu111 \
+    torchaudio==0.10.0 \
+    -f https://download.pytorch.org/whl/torch_stable.html
+
+and it throws the following error when running ``tdnn/train.py``:
+
+.. code-block::
+
+  OSError: libtorch_hip.so: cannot open shared object file: no such file or directory
+
+The fix is to specify the CUDA version while installing ``torchaudio``. That
+is, change ``torchaudio==0.10.0`` to ``torchaudio==0.10.0+cu11```. Therefore,
+the correct command is:
+
+.. code-block:: bash
+
+  pip install \
+    torch==1.10.0+cu111 \
+    torchvision==0.11.0+cu111 \
+    torchaudio==0.10.0+cu111 \
+    -f https://download.pytorch.org/whl/torch_stable.html
+
+AttributeError: module 'distutils' has no attribute 'version'
+-------------------------------------------------------------
+
+The error log is:
+
+.. code-block::
+
+  Traceback (most recent call last):
+    File "./tdnn/train.py", line 14, in <module>
+      from asr_datamodule import YesNoAsrDataModule
+    File "/home/xxx/code/next-gen-kaldi/icefall/egs/yesno/ASR/tdnn/asr_datamodule.py", line 34, in <module>
+      from icefall.dataset.datamodule import DataModule
+    File "/home/xxx/code/next-gen-kaldi/icefall/icefall/__init__.py", line 3, in <module>
+      from . import (
+    File "/home/xxx/code/next-gen-kaldi/icefall/icefall/decode.py", line 23, in <module>
+      from icefall.utils import add_eos, add_sos, get_texts
+    File "/home/xxx/code/next-gen-kaldi/icefall/icefall/utils.py", line 39, in <module>
+      from torch.utils.tensorboard import SummaryWriter
+    File "/home/xxx/tool/miniconda3/envs/yyy/lib/python3.8/site-packages/torch/utils/tensorboard/__init__.py", line 4, in <module>
+      LooseVersion = distutils.version.LooseVersion
+  AttributeError: module 'distutils' has no attribute 'version'
+
+The fix is:
+
+.. code-block:: bash
+
+  pip uninstall setuptools
+
+  pip install setuptools==58.0.4
+
+ImportError: libpython3.10.so.1.0: cannot open shared object file: No such file or directory
+--------------------------------------------------------------------------------------------
+
+If you are using ``conda`` and encounter the following issue:
+
+.. code-block::
+
+  Traceback (most recent call last):
+    File "/k2-dev/yangyifan/anaconda3/envs/icefall/lib/python3.10/site-packages/k2-1.23.3.dev20230112+cuda11.6.torch1.13.1-py3.10-linux-x86_64.egg/k2/__init__.py", line 24, in <module>
+      from _k2 import DeterminizeWeightPushingType
+  ImportError: libpython3.10.so.1.0: cannot open shared object file: No such file or directory
+
+  During handling of the above exception, another exception occurred:
+
+  Traceback (most recent call last):
+    File "/k2-dev/yangyifan/icefall/egs/librispeech/ASR/./pruned_transducer_stateless7_ctc_bs/decode.py", line 104, in <module>
+      import k2
+    File "/k2-dev/yangyifan/anaconda3/envs/icefall/lib/python3.10/site-packages/k2-1.23.3.dev20230112+cuda11.6.torch1.13.1-py3.10-linux-x86_64.egg/k2/__init__.py", line 30, in <module>
+      raise ImportError(
+  ImportError: libpython3.10.so.1.0: cannot open shared object file: No such file or directory
+  Note: If you're using anaconda and importing k2 on MacOS,
+        you can probably fix this by setting the environment variable:
+    export DYLD_LIBRARY_PATH=$CONDA_PREFIX/lib/python3.10/site-packages:$DYLD_LIBRARY_PATH
+
+Please first try to find where ``libpython3.10.so.1.0`` locates.
+
+For instance,
+
+.. code-block:: bash
+
+  cd $CONDA_PREFIX/lib
+  find . -name "libpython*"
+
+If you are able to find it inside ``$CODNA_PREFIX/lib``, please set the
+following environment variable:
+
+.. code-block:: bash
+
+  export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH

docs/source/index.rst

@@ -21,7 +21,16 @@ speech recognition recipes using `k2 <https://github.com/k2-fsa/k2>`_.
    :caption: Contents:
 
    installation/index
+   faqs
    model-export/index
+
+.. toctree::
+   :maxdepth: 3
+
    recipes/index
+
+.. toctree::
+   :maxdepth: 2
+
    contributing/index
    huggingface/index

docs/source/installation/index.rst

@@ -393,6 +393,17 @@ Now let us run the training part:
   We use ``export CUDA_VISIBLE_DEVICES=""`` so that ``icefall`` uses CPU
   even if there are GPUs available.
 
+.. hint::
+
+   In case you get a ``Segmentation fault (core dump)`` error, please use:
+
+      .. code-block:: bash
+
+        export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+   See more at `<https://github.com/k2-fsa/icefall/issues/674>` if you are
+   interested.
+
 The training log is given below:
 
 .. code-block::

docs/source/model-export/code/export-conv-emformer-transducer-for-ncnn-output.txt

@@ -0,0 +1,21 @@
+2023-01-11 12:15:38,677 INFO [export-for-ncnn.py:220] device: cpu
+2023-01-11 12:15:38,681 INFO [export-for-ncnn.py:229] {'best_train_loss': inf, 'best_valid_loss': inf, 'best_train_epoch': -1, 'best_v
+alid_epoch': -1, 'batch_idx_train': 0, 'log_interval': 50, 'reset_interval': 200, 'valid_interval': 3000, 'feature_dim': 80, 'subsampl
+ing_factor': 4, 'decoder_dim': 512, 'joiner_dim': 512, 'model_warm_step': 3000, 'env_info': {'k2-version': '1.23.2', 'k2-build-type':
+'Release', 'k2-with-cuda': True, 'k2-git-sha1': 'a34171ed85605b0926eebbd0463d059431f4f74a', 'k2-git-date': 'Wed Dec 14 00:06:38 2022',
+ 'lhotse-version': '1.12.0.dev+missing.version.file', 'torch-version': '1.10.0+cu102', 'torch-cuda-available': False, 'torch-cuda-vers
+ion': '10.2', 'python-version': '3.8', 'icefall-git-branch': 'fix-stateless3-train-2022-12-27', 'icefall-git-sha1': '530e8a1-dirty', '
+icefall-git-date': 'Tue Dec 27 13:59:18 2022', 'icefall-path': '/star-fj/fangjun/open-source/icefall', 'k2-path': '/star-fj/fangjun/op
+en-source/k2/k2/python/k2/__init__.py', 'lhotse-path': '/star-fj/fangjun/open-source/lhotse/lhotse/__init__.py', 'hostname': 'de-74279
+-k2-train-3-1220120619-7695ff496b-s9n4w', 'IP address': '127.0.0.1'}, 'epoch': 30, 'iter': 0, 'avg': 1, 'exp_dir': PosixPath('icefa
+ll-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp'), 'bpe_model': './icefall-asr-librispeech-conv-emformer-transdu
+cer-stateless2-2022-07-05//data/lang_bpe_500/bpe.model', 'jit': False, 'context_size': 2, 'use_averaged_model': False, 'encoder_dim':
+512, 'nhead': 8, 'dim_feedforward': 2048, 'num_encoder_layers': 12, 'cnn_module_kernel': 31, 'left_context_length': 32, 'chunk_length'
+: 32, 'right_context_length': 8, 'memory_size': 32, 'blank_id': 0, 'vocab_size': 500}
+2023-01-11 12:15:38,681 INFO [export-for-ncnn.py:231] About to create model
+2023-01-11 12:15:40,053 INFO [checkpoint.py:112] Loading checkpoint from icefall-asr-librispeech-conv-emformer-transducer-stateless2-2
+022-07-05/exp/epoch-30.pt
+2023-01-11 12:15:40,708 INFO [export-for-ncnn.py:315] Number of model parameters: 75490012
+2023-01-11 12:15:41,681 INFO [export-for-ncnn.py:318] Using torch.jit.trace()
+2023-01-11 12:15:41,681 INFO [export-for-ncnn.py:320] Exporting encoder
+2023-01-11 12:15:41,682 INFO [export-for-ncnn.py:149] chunk_length: 32, right_context_length: 8

docs/source/model-export/code/generate-int-8-scale-table-for-conv-emformer.txt

@@ -0,0 +1,104 @@
+Don't Use GPU. has_gpu: 0, config.use_vulkan_compute: 1
+num encoder conv layers: 88
+num joiner conv layers: 3
+num files: 3
+Processing ../test_wavs/1089-134686-0001.wav
+Processing ../test_wavs/1221-135766-0001.wav
+Processing ../test_wavs/1221-135766-0002.wav
+Processing ../test_wavs/1089-134686-0001.wav
+Processing ../test_wavs/1221-135766-0001.wav
+Processing ../test_wavs/1221-135766-0002.wav
+----------encoder----------
+conv_87                                  : max = 15.942385        threshold = 15.938493        scale = 7.968131
+conv_88                                  : max = 35.442448        threshold = 15.549335        scale = 8.167552
+conv_89                                  : max = 23.228289        threshold = 8.001738         scale = 15.871552
+linear_90                                : max = 3.976146         threshold = 1.101789         scale = 115.267128
+linear_91                                : max = 6.962030         threshold = 5.162033         scale = 24.602713
+linear_92                                : max = 12.323041        threshold = 3.853959         scale = 32.953129
+linear_94                                : max = 6.905416         threshold = 4.648006         scale = 27.323545
+linear_93                                : max = 6.905416         threshold = 5.474093         scale = 23.200188
+linear_95                                : max = 1.888012         threshold = 1.403563         scale = 90.483986
+linear_96                                : max = 6.856741         threshold = 5.398679         scale = 23.524273
+linear_97                                : max = 9.635942         threshold = 2.613655         scale = 48.590950
+linear_98                                : max = 6.460340         threshold = 5.670146         scale = 22.398010
+linear_99                                : max = 9.532276         threshold = 2.585537         scale = 49.119396
+linear_101                               : max = 6.585871         threshold = 5.719224         scale = 22.205809
+linear_100                               : max = 6.585871         threshold = 5.751382         scale = 22.081648
+linear_102                               : max = 1.593344         threshold = 1.450581         scale = 87.551147
+linear_103                               : max = 6.592681         threshold = 5.705824         scale = 22.257959
+linear_104                               : max = 8.752957         threshold = 1.980955         scale = 64.110489
+linear_105                               : max = 6.696240         threshold = 5.877193         scale = 21.608953
+linear_106                               : max = 9.059659         threshold = 2.643138         scale = 48.048950
+linear_108                               : max = 6.975461         threshold = 4.589567         scale = 27.671457
+linear_107                               : max = 6.975461         threshold = 6.190381         scale = 20.515701
+linear_109                               : max = 3.710759         threshold = 2.305635         scale = 55.082436
+linear_110                               : max = 7.531228         threshold = 5.731162         scale = 22.159557
+linear_111                               : max = 10.528083        threshold = 2.259322         scale = 56.211544
+linear_112                               : max = 8.148807         threshold = 5.500842         scale = 23.087374
+linear_113                               : max = 8.592566         threshold = 1.948851         scale = 65.166611
+linear_115                               : max = 8.437109         threshold = 5.608947         scale = 22.642395
+linear_114                               : max = 8.437109         threshold = 6.193942         scale = 20.503904
+linear_116                               : max = 3.966980         threshold = 3.200896         scale = 39.676392
+linear_117                               : max = 9.451303         threshold = 6.061664         scale = 20.951344
+linear_118                               : max = 12.077262        threshold = 3.965800         scale = 32.023804
+linear_119                               : max = 9.671615         threshold = 4.847613         scale = 26.198460
+linear_120                               : max = 8.625638         threshold = 3.131427         scale = 40.556595
+linear_122                               : max = 10.274080        threshold = 4.888716         scale = 25.978189
+linear_121                               : max = 10.274080        threshold = 5.420480         scale = 23.429659
+linear_123                               : max = 4.826197         threshold = 3.599617         scale = 35.281532
+linear_124                               : max = 11.396383        threshold = 7.325849         scale = 17.335875
+linear_125                               : max = 9.337198         threshold = 3.941410         scale = 32.221970
+linear_126                               : max = 9.699965         threshold = 4.842878         scale = 26.224073
+linear_127                               : max = 8.775370         threshold = 3.884215         scale = 32.696438
+linear_129                               : max = 9.872276         threshold = 4.837319         scale = 26.254213
+linear_128                               : max = 9.872276         threshold = 7.180057         scale = 17.687883
+linear_130                               : max = 4.150427         threshold = 3.454298         scale = 36.765789
+linear_131                               : max = 11.112692        threshold = 7.924847         scale = 16.025545
+linear_132                               : max = 11.852893        threshold = 3.116593         scale = 40.749626
+linear_133                               : max = 11.517084        threshold = 5.024665         scale = 25.275314
+linear_134                               : max = 10.683807        threshold = 3.878618         scale = 32.743618
+linear_136                               : max = 12.421055        threshold = 6.322729         scale = 20.086264
+linear_135                               : max = 12.421055        threshold = 5.309880         scale = 23.917679
+linear_137                               : max = 4.827781         threshold = 3.744595         scale = 33.915554
+linear_138                               : max = 14.422395        threshold = 7.742882         scale = 16.402161
+linear_139                               : max = 8.527538         threshold = 3.866123         scale = 32.849449
+linear_140                               : max = 12.128619        threshold = 4.657793         scale = 27.266134
+linear_141                               : max = 9.839593         threshold = 3.845993         scale = 33.021378
+linear_143                               : max = 12.442304        threshold = 7.099039         scale = 17.889746
+linear_142                               : max = 12.442304        threshold = 5.325038         scale = 23.849592
+linear_144                               : max = 5.929444         threshold = 5.618206         scale = 22.605080
+linear_145                               : max = 13.382126        threshold = 9.321095         scale = 13.625010
+linear_146                               : max = 9.894987         threshold = 3.867645         scale = 32.836517
+linear_147                               : max = 10.915313        threshold = 4.906028         scale = 25.886522
+linear_148                               : max = 9.614287         threshold = 3.908151         scale = 32.496181
+linear_150                               : max = 11.724932        threshold = 4.485588         scale = 28.312899
+linear_149                               : max = 11.724932        threshold = 5.161146         scale = 24.606939
+linear_151                               : max = 7.164453         threshold = 5.847355         scale = 21.719223
+linear_152                               : max = 13.086471        threshold = 5.984121         scale = 21.222834
+linear_153                               : max = 11.099524        threshold = 3.991601         scale = 31.816805
+linear_154                               : max = 10.054585        threshold = 4.489706         scale = 28.286930
+linear_155                               : max = 12.389185        threshold = 3.100321         scale = 40.963501
+linear_157                               : max = 9.982999         threshold = 5.154796         scale = 24.637253
+linear_156                               : max = 9.982999         threshold = 8.537706         scale = 14.875190
+linear_158                               : max = 8.420287         threshold = 6.502287         scale = 19.531588
+linear_159                               : max = 25.014746        threshold = 9.423280         scale = 13.477261
+linear_160                               : max = 45.633553        threshold = 5.715335         scale = 22.220921
+linear_161                               : max = 20.371849        threshold = 5.117830         scale = 24.815203
+linear_162                               : max = 12.492933        threshold = 3.126283         scale = 40.623318
+linear_164                               : max = 20.697504        threshold = 4.825712         scale = 26.317358
+linear_163                               : max = 20.697504        threshold = 5.078367         scale = 25.008038
+linear_165                               : max = 9.023975         threshold = 6.836278         scale = 18.577358
+linear_166                               : max = 34.860619        threshold = 7.259792         scale = 17.493614
+linear_167                               : max = 30.380934        threshold = 5.496160         scale = 23.107042
+linear_168                               : max = 20.691216        threshold = 4.733317         scale = 26.831076
+linear_169                               : max = 9.723948         threshold = 3.952728         scale = 32.129707
+linear_171                               : max = 21.034811        threshold = 5.366547         scale = 23.665123
+linear_170                               : max = 21.034811        threshold = 5.356277         scale = 23.710501
+linear_172                               : max = 10.556884        threshold = 5.729481         scale = 22.166058
+linear_173                               : max = 20.033039        threshold = 10.207264        scale = 12.442120
+linear_174                               : max = 11.597379        threshold = 2.658676         scale = 47.768131
+----------joiner----------
+linear_2                                 : max = 19.293503        threshold = 14.305265        scale = 8.877850
+linear_1                                 : max = 10.812222        threshold = 8.766452         scale = 14.487047
+linear_3                                 : max = 0.999999         threshold = 0.999755         scale = 127.031174
+ncnn int8 calibration table create success, best wish for your int8 inference has a low accuracy loss...\(^0^)/...233...

docs/source/model-export/code/test-stremaing-ncnn-decode-conv-emformer-transducer-libri.txt

@@ -0,0 +1,7 @@
+2023-01-11 14:02:12,216 INFO [streaming-ncnn-decode.py:320] {'tokens': './icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/data/lang_bpe_500/tokens.txt', 'encoder_param_filename': './icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.param', 'encoder_bin_filename': './icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.bin', 'decoder_param_filename': './icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.param', 'decoder_bin_filename': './icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.bin', 'joiner_param_filename': './icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.param', 'joiner_bin_filename': './icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.bin', 'sound_filename': './icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/test_wavs/1089-134686-0001.wav'}
+T 51 32
+2023-01-11 14:02:13,141 INFO [streaming-ncnn-decode.py:328] Constructing Fbank computer
+2023-01-11 14:02:13,151 INFO [streaming-ncnn-decode.py:331] Reading sound files: ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/test_wavs/1089-134686-0001.wav
+2023-01-11 14:02:13,176 INFO [streaming-ncnn-decode.py:336] torch.Size([106000])
+2023-01-11 14:02:17,581 INFO [streaming-ncnn-decode.py:380] ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/test_wavs/1089-134686-0001.wav
+2023-01-11 14:02:17,581 INFO [streaming-ncnn-decode.py:381] AFTER EARLY NIGHTFALL THE YELLOW LAMPS WOULD LIGHT UP HERE AND THERE THE SQUALID QUARTER OF THE BROTHELS

docs/source/model-export/export-ncnn.rst

@@ -1,12 +1,771 @@
 Export to ncnn
 ==============
 
-We support exporting LSTM transducer models to `ncnn <https://github.com/tencent/ncnn>`_.
-
-Please refer to :ref:`export-model-for-ncnn` for details.
+We support exporting both
+`LSTM transducer models <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/lstm_transducer_stateless2>`_
+and
+`ConvEmformer transducer models <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/conv_emformer_transducer_stateless2>`_
+to `ncnn <https://github.com/tencent/ncnn>`_.
 
 We also provide `<https://github.com/k2-fsa/sherpa-ncnn>`_
 performing speech recognition using ``ncnn`` with exported models.
-It has been tested on Linux, macOS, Windows, and Raspberry Pi. The project is
-self-contained and can be statically linked to produce a binary containing
-everything needed.
+It has been tested on Linux, macOS, Windows, ``Android``, and ``Raspberry Pi``.
+
+`sherpa-ncnn`_ is self-contained and can be statically linked to produce
+a binary containing everything needed. Please refer
+to its documentation for details:
+
+ - `<https://k2-fsa.github.io/sherpa/ncnn/index.html>`_
+
+
+Export LSTM transducer models
+-----------------------------
+
+Please refer to :ref:`export-lstm-transducer-model-for-ncnn` for details.
+
+
+
+Export ConvEmformer transducer models
+-------------------------------------
+
+We use the pre-trained model from the following repository as an example:
+
+  - `<https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05>`_
+
+We will show you step by step how to export it to `ncnn`_ and run it with `sherpa-ncnn`_.
+
+.. hint::
+
+  We use ``Ubuntu 18.04``, ``torch 1.10``, and ``Python 3.8`` for testing.
+
+.. caution::
+
+  Please use a more recent version of PyTorch. For instance, ``torch 1.8``
+  may ``not`` work.
+
+1. Download the pre-trained model
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. hint::
+
+  You can also refer to `<https://k2-fsa.github.io/sherpa/cpp/pretrained_models/online_transducer.html#icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05>`_ to download the pre-trained model.
+
+  You have to install `git-lfs`_ before you continue.
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+
+  GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05
+
+  git lfs pull --include "exp/pretrained-epoch-30-avg-10-averaged.pt"
+  git lfs pull --include "data/lang_bpe_500/bpe.model"
+
+  cd ..
+
+.. note::
+
+  We download ``exp/pretrained-xxx.pt``, not ``exp/cpu-jit_xxx.pt``.
+
+
+In the above code, we download the pre-trained model into the directory
+``egs/librispeech/ASR/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05``.
+
+2. Install ncnn and pnnx
+^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. code-block:: bash
+
+  # We put ncnn into $HOME/open-source/ncnn
+  # You can change it to anywhere you like
+
+  cd $HOME
+  mkdir -p open-source
+  cd open-source
+
+  git clone https://github.com/csukuangfj/ncnn
+  cd ncnn
+  git submodule update --recursive --init
+
+  # Note: We don't use "python setup.py install" or "pip install ." here
+
+  mkdir -p build-wheel
+  cd build-wheel
+
+  cmake \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DNCNN_PYTHON=ON \
+    -DNCNN_BUILD_BENCHMARK=OFF \
+    -DNCNN_BUILD_EXAMPLES=OFF \
+    -DNCNN_BUILD_TOOLS=ON \
+  ..
+
+  make -j4
+
+  cd ..
+
+  # Note: $PWD here is $HOME/open-source/ncnn
+
+  export PYTHONPATH=$PWD/python:$PYTHONPATH
+  export PATH=$PWD/tools/pnnx/build/src:$PATH
+  export PATH=$PWD/build-wheel/tools/quantize:$PATH
+
+  # Now build pnnx
+  cd tools/pnnx
+  mkdir build
+  cd build
+  cmake ..
+  make -j4
+
+  ./src/pnnx
+
+Congratulations! You have successfully installed the following components:
+
+  - ``pnxx``, which is an executable located in
+    ``$HOME/open-source/ncnn/tools/pnnx/build/src``. We will use
+    it to convert models exported by ``torch.jit.trace()``.
+  - ``ncnn2int8``, which is an executable located in
+    ``$HOME/open-source/ncnn/build-wheel/tools/quantize``. We will use
+    it to quantize our models to ``int8``.
+  - ``ncnn.cpython-38-x86_64-linux-gnu.so``, which is a Python module located
+    in ``$HOME/open-source/ncnn/python/ncnn``.
+
+    .. note::
+
+      I am using ``Python 3.8``, so it
+      is ``ncnn.cpython-38-x86_64-linux-gnu.so``. If you use a different
+      version, say, ``Python 3.9``, the name would be
+      ``ncnn.cpython-39-x86_64-linux-gnu.so``.
+
+      Also, if you are not using Linux, the file name would also be different.
+      But that does not matter. As long as you can compile it, it should work.
+
+We have set up ``PYTHONPATH`` so that you can use ``import ncnn`` in your
+Python code. We have also set up ``PATH`` so that you can use
+``pnnx`` and ``ncnn2int8`` later in your terminal.
+
+.. caution::
+
+  Please don't use `<https://github.com/tencent/ncnn>`_.
+  We have made some modifications to the offical `ncnn`_.
+
+  We will synchronize `<https://github.com/csukuangfj/ncnn>`_ periodically
+  with the official one.
+
+3. Export the model via torch.jit.trace()
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+First, let us rename our pre-trained model:
+
+.. code-block::
+
+  cd egs/librispeech/ASR
+
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp
+
+  ln -s pretrained-epoch-30-avg-10-averaged.pt epoch-30.pt
+
+  cd ../..
+
+Next, we use the following code to export our model:
+
+.. code-block:: bash
+
+  dir=./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/
+
+  ./conv_emformer_transducer_stateless2/export-for-ncnn.py \
+    --exp-dir $dir/exp \
+    --bpe-model $dir/data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 1 \
+    --use-averaged-model 0 \
+    \
+    --num-encoder-layers 12 \
+    --chunk-length 32 \
+    --cnn-module-kernel 31 \
+    --left-context-length 32 \
+    --right-context-length 8 \
+    --memory-size 32 \
+    --encoder-dim 512
+
+.. hint::
+
+  We have renamed our model to ``epoch-30.pt`` so that we can use ``--epoch 30``.
+  There is only one pre-trained model, so we use ``--avg 1 --use-averaged-model 0``.
+
+  If you have trained a model by yourself and if you have all checkpoints
+  available, please first use ``decode.py`` to tune ``--epoch --avg``
+  and select the best combination with with ``--use-averaged-model 1``.
+
+.. note::
+
+  You will see the following log output:
+
+  .. literalinclude:: ./code/export-conv-emformer-transducer-for-ncnn-output.txt
+
+  The log shows the model has ``75490012`` parameters, i.e., ``~75 M``.
+
+  .. code-block::
+
+    ls -lh icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/pretrained-epoch-30-avg-10-averaged.pt
+
+    -rw-r--r-- 1 kuangfangjun root 289M Jan 11 12:05 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/pretrained-epoch-30-avg-10-averaged.pt
+
+  You can see that the file size of the pre-trained model is ``289 MB``, which
+  is roughly ``75490012*4/1024/1024 = 287.97 MB``.
+
+After running ``conv_emformer_transducer_stateless2/export-for-ncnn.py``,
+we will get the following files:
+
+.. code-block:: bash
+
+  ls -lh icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/*pnnx*
+
+  -rw-r--r-- 1 kuangfangjun root 1010K Jan 11 12:15 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.pt
+  -rw-r--r-- 1 kuangfangjun root  283M Jan 11 12:15 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.pt
+  -rw-r--r-- 1 kuangfangjun root  3.0M Jan 11 12:15 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.pt
+
+
+.. _conv-emformer-step-3-export-torchscript-model-via-pnnx:
+
+3. Export torchscript model via pnnx
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. hint::
+
+  Make sure you have set up the ``PATH`` environment variable. Otherwise,
+  it will throw an error saying that ``pnnx`` could not be found.
+
+Now, it's time to export our models to `ncnn`_ via ``pnnx``.
+
+.. code-block::
+
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+  pnnx ./encoder_jit_trace-pnnx.pt
+  pnnx ./decoder_jit_trace-pnnx.pt
+  pnnx ./joiner_jit_trace-pnnx.pt
+
+It will generate the following files:
+
+.. code-block:: bash
+
+  ls -lh  icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/*ncnn*{bin,param}
+
+  -rw-r--r-- 1 kuangfangjun root 503K Jan 11 12:38 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  437 Jan 11 12:38 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 142M Jan 11 12:36 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  79K Jan 11 12:36 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 1.5M Jan 11 12:38 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  488 Jan 11 12:38 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.param
+
+There are two types of files:
+
+- ``param``: It is a text file containing the model architectures. You can
+  use a text editor to view its content.
+- ``bin``: It is a binary file containing the model parameters.
+
+We compare the file sizes of the models below before and after converting via ``pnnx``:
+
+.. see https://tableconvert.com/restructuredtext-generator
+
++----------------------------------+------------+
+| File name                        | File size  |
++==================================+============+
+| encoder_jit_trace-pnnx.pt        | 283 MB     |
++----------------------------------+------------+
+| decoder_jit_trace-pnnx.pt        | 1010 KB    |
++----------------------------------+------------+
+| joiner_jit_trace-pnnx.pt         | 3.0 MB     |
++----------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin  | 142 MB     |
++----------------------------------+------------+
+| decoder_jit_trace-pnnx.ncnn.bin  | 503 KB     |
++----------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin   | 1.5 MB     |
++----------------------------------+------------+
+
+You can see that the file sizes of the models after conversion are about one half
+of the models before conversion:
+
+  - encoder: 283 MB vs 142 MB
+  - decoder: 1010 KB vs 503 KB
+  - joiner: 3.0 MB vs 1.5 MB
+
+The reason is that by default ``pnnx`` converts ``float32`` parameters
+to ``float16``. A ``float32`` parameter occupies 4 bytes, while it is 2 bytes
+for ``float16``. Thus, it is ``twice smaller`` after conversion.
+
+.. hint::
+
+  If you use ``pnnx ./encoder_jit_trace-pnnx.pt fp16=0``, then ``pnnx``
+  won't convert ``float32`` to ``float16``.
+
+4. Test the exported models in icefall
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. note::
+
+  We assume you have set up the environment variable ``PYTHONPATH`` when
+  building `ncnn`_.
+
+Now we have successfully converted our pre-trained model to `ncnn`_ format.
+The generated 6 files are what we need. You can use the following code to
+test the converted models:
+
+.. code-block:: bash
+
+  ./conv_emformer_transducer_stateless2/streaming-ncnn-decode.py \
+    --tokens ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/data/lang_bpe_500/tokens.txt \
+    --encoder-param-filename ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.param \
+    --encoder-bin-filename ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.bin \
+    --decoder-param-filename ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.param \
+    --decoder-bin-filename ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.bin \
+    --joiner-param-filename ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.param \
+    --joiner-bin-filename ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.bin \
+    ./icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/test_wavs/1089-134686-0001.wav
+
+.. hint::
+
+  `ncnn`_ supports only ``batch size == 1``, so ``streaming-ncnn-decode.py`` accepts
+  only 1 wave file as input.
+
+The output is given below:
+
+.. literalinclude:: ./code/test-stremaing-ncnn-decode-conv-emformer-transducer-libri.txt
+
+Congratulations! You have successfully exported a model from PyTorch to `ncnn`_!
+
+
+.. _conv-emformer-modify-the-exported-encoder-for-sherpa-ncnn:
+
+5. Modify the exported encoder for sherpa-ncnn
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+In order to use the exported models in `sherpa-ncnn`_, we have to modify
+``encoder_jit_trace-pnnx.ncnn.param``.
+
+Let us have a look at the first few lines of ``encoder_jit_trace-pnnx.ncnn.param``:
+
+.. code-block::
+
+  7767517
+  1060 1342
+  Input                    in0                      0 1 in0
+
+**Explanation** of the above three lines:
+
+  1. ``7767517``, it is a magic number and should not be changed.
+  2. ``1060 1342``, the first number ``1060`` specifies the number of layers
+     in this file, while ``1342`` specifies the number of intermediate outputs
+     of this file
+  3. ``Input in0 0 1 in0``, ``Input`` is the layer type of this layer; ``in0``
+     is the layer name of this layer; ``0`` means this layer has no input;
+     ``1`` means this layer has one output; ``in0`` is the output name of
+     this layer.
+
+We need to add 1 extra line and also increment the number of layers.
+The result looks like below:
+
+.. code-block:: bash
+
+  7767517
+  1061 1342
+  SherpaMetaData           sherpa_meta_data1        0 0 0=1 1=12 2=32 3=31 4=8 5=32 6=8 7=512
+  Input                    in0                      0 1 in0
+
+**Explanation**
+
+  1. ``7767517``, it is still the same
+  2. ``1061 1342``, we have added an extra layer, so we need to update ``1060`` to ``1061``.
+     We don't need to change ``1342`` since the newly added layer has no inputs or outputs.
+  3. ``SherpaMetaData  sherpa_meta_data1  0 0 0=1 1=12 2=32 3=31 4=8 5=32 6=8 7=512``
+     This line is newly added. Its explanation is given below:
+
+      - ``SherpaMetaData`` is the type of this layer. Must be ``SherpaMetaData``.
+      - ``sherpa_meta_data1`` is the name of this layer. Must be ``sherpa_meta_data1``.
+      - ``0 0`` means this layer has no inputs or output. Must be ``0 0``
+      - ``0=1``, 0 is the key and 1 is the value. MUST be ``0=1``
+      - ``1=12``, 1 is the key and 12 is the value of the
+        parameter ``--num-encoder-layers`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+      - ``2=32``, 2 is the key and 32 is the value of the
+        parameter ``--memory-size`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+      - ``3=31``, 3 is the key and 31 is the value of the
+        parameter ``--cnn-module-kernel`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+      - ``4=8``, 4 is the key and 8 is the value of the
+        parameter ``--left-context-length`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+      - ``5=32``, 5 is the key and 32 is the value of the
+        parameter ``--chunk-length`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+      - ``6=8``, 6 is the key and 8 is the value of the
+        parameter ``--right-context-length`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+      - ``7=512``, 7 is the key and 512 is the value of the
+        parameter ``--encoder-dim`` that you provided when running
+        ``conv_emformer_transducer_stateless2/export-for-ncnn.py``.
+
+      For ease of reference, we list the key-value pairs that you need to add
+      in the following table. If your model has a different setting, please
+      change the values for ``SherpaMetaData`` accordingly. Otherwise, you
+      will be ``SAD``.
+
+          +------+-----------------------------+
+          | key  | value                       |
+          +======+=============================+
+          | 0    | 1 (fixed)                   |
+          +------+-----------------------------+
+          | 1    | ``--num-encoder-layers``    |
+          +------+-----------------------------+
+          | 2    | ``--memory-size``           |
+          +------+-----------------------------+
+          | 3    | ``--cnn-module-kernel``     |
+          +------+-----------------------------+
+          | 4    | ``--left-context-length``   |
+          +------+-----------------------------+
+          | 5    | ``--chunk-length``          |
+          +------+-----------------------------+
+          | 6    | ``--right-context-length``  |
+          +------+-----------------------------+
+          | 7    | ``--encoder-dim``           |
+          +------+-----------------------------+
+
+  4. ``Input in0 0 1 in0``. No need to change it.
+
+.. caution::
+
+  When you add a new layer ``SherpaMetaData``, please remember to update the
+  number of layers. In our case, update  ``1060`` to ``1061``. Otherwise,
+  you will be SAD later.
+
+.. hint::
+
+  After adding the new layer ``SherpaMetaData``, you cannot use this model
+  with ``streaming-ncnn-decode.py`` anymore since ``SherpaMetaData`` is
+  supported only in `sherpa-ncnn`_.
+
+.. hint::
+
+  `ncnn`_ is very flexible. You can add new layers to it just by text-editing
+  the ``param`` file! You don't need to change the ``bin`` file.
+
+Now you can use this model in `sherpa-ncnn`_.
+Please refer to the following documentation:
+
+  - Linux/macOS/Windows/arm/aarch64: `<https://k2-fsa.github.io/sherpa/ncnn/install/index.html>`_
+  - Android: `<https://k2-fsa.github.io/sherpa/ncnn/android/index.html>`_
+  - Python: `<https://k2-fsa.github.io/sherpa/ncnn/python/index.html>`_
+
+We have a list of pre-trained models that have been exported for `sherpa-ncnn`_:
+
+  - `<https://k2-fsa.github.io/sherpa/ncnn/pretrained_models/index.html>`_
+
+    You can find more usages there.
+
+6. (Optional) int8 quantization with sherpa-ncnn
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+This step is optional.
+
+In this step, we describe how to quantize our model with ``int8``.
+
+Change :ref:`conv-emformer-step-3-export-torchscript-model-via-pnnx` to
+disable ``fp16`` when using ``pnnx``:
+
+.. code-block::
+
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+  pnnx ./encoder_jit_trace-pnnx.pt fp16=0
+  pnnx ./decoder_jit_trace-pnnx.pt
+  pnnx ./joiner_jit_trace-pnnx.pt fp16=0
+
+.. note::
+
+  We add ``fp16=0`` when exporting the encoder and joiner. `ncnn`_ does not
+  support quantizing the decoder model yet. We will update this documentation
+  once `ncnn`_ supports it. (Maybe in this year, 2023).
+
+It will generate the following files
+
+.. code-block:: bash
+
+  ls -lh icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/*_jit_trace-pnnx.ncnn.{param,bin}
+
+  -rw-r--r-- 1 kuangfangjun root 503K Jan 11 15:56 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  437 Jan 11 15:56 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/decoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 283M Jan 11 15:56 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  79K Jan 11 15:56 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/encoder_jit_trace-pnnx.ncnn.param
+  -rw-r--r-- 1 kuangfangjun root 3.0M Jan 11 15:56 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.bin
+  -rw-r--r-- 1 kuangfangjun root  488 Jan 11 15:56 icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/joiner_jit_trace-pnnx.ncnn.param
+
+Let us compare again the file sizes:
+
++----------------------------------------+------------+
+| File name                              | File size  |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.pt              | 283 MB     |
++----------------------------------------+------------+
+| decoder_jit_trace-pnnx.pt              | 1010 KB    |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.pt               | 3.0 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin (fp16) | 142 MB     |
++----------------------------------------+------------+
+| decoder_jit_trace-pnnx.ncnn.bin (fp16) | 503 KB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin  (fp16) | 1.5 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin (fp32) | 283 MB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin  (fp32) | 3.0 MB     |
++----------------------------------------+------------+
+
+You can see that the file sizes are doubled when we disable ``fp16``.
+
+.. note::
+
+  You can again use ``streaming-ncnn-decode.py`` to test the exported models.
+
+Next, follow :ref:`conv-emformer-modify-the-exported-encoder-for-sherpa-ncnn`
+to modify ``encoder_jit_trace-pnnx.ncnn.param``.
+
+Change
+
+.. code-block:: bash
+
+  7767517
+  1060 1342
+  Input                    in0                      0 1 in0
+
+to
+
+.. code-block:: bash
+
+  7767517
+  1061 1342
+  SherpaMetaData           sherpa_meta_data1        0 0 0=1 1=12 2=32 3=31 4=8 5=32 6=8 7=512
+  Input                    in0                      0 1 in0
+
+.. caution::
+
+  Please follow :ref:`conv-emformer-modify-the-exported-encoder-for-sherpa-ncnn`
+  to change the values for ``SherpaMetaData`` if your model uses a different setting.
+
+
+Next, let us compile `sherpa-ncnn`_ since we will quantize our models within
+`sherpa-ncnn`_.
+
+.. code-block:: bash
+
+  # We will download sherpa-ncnn to $HOME/open-source/
+  # You can change it to anywhere you like.
+  cd $HOME
+  mkdir -p open-source
+
+  cd open-source
+  git clone https://github.com/k2-fsa/sherpa-ncnn
+  cd sherpa-ncnn
+  mkdir build
+  cd build
+  cmake ..
+  make -j 4
+
+  ./bin/generate-int8-scale-table
+
+  export PATH=$HOME/open-source/sherpa-ncnn/build/bin:$PATH
+
+The output of the above commands are:
+
+.. code-block:: bash
+
+  (py38) kuangfangjun:build$ generate-int8-scale-table
+  Please provide 10 arg. Currently given: 1
+  Usage:
+  generate-int8-scale-table encoder.param encoder.bin decoder.param decoder.bin joiner.param joiner.bin encoder-scale-table.txt joiner-scale-table.txt wave_filenames.txt
+
+  Each line in wave_filenames.txt is a path to some 16k Hz mono wave file.
+
+We need to create a file ``wave_filenames.txt``, in which we need to put
+some calibration wave files. For testing purpose, we put the ``test_wavs``
+from the pre-trained model repository `<https://huggingface.co/Zengwei/icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05>`_
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+  cat <<EOF > wave_filenames.txt
+  ../test_wavs/1089-134686-0001.wav
+  ../test_wavs/1221-135766-0001.wav
+  ../test_wavs/1221-135766-0002.wav
+  EOF
+
+Now we can calculate the scales needed for quantization with the calibration data:
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+  generate-int8-scale-table \
+    ./encoder_jit_trace-pnnx.ncnn.param \
+    ./encoder_jit_trace-pnnx.ncnn.bin \
+    ./decoder_jit_trace-pnnx.ncnn.param \
+    ./decoder_jit_trace-pnnx.ncnn.bin \
+    ./joiner_jit_trace-pnnx.ncnn.param \
+    ./joiner_jit_trace-pnnx.ncnn.bin \
+    ./encoder-scale-table.txt \
+    ./joiner-scale-table.txt \
+    ./wave_filenames.txt
+
+The output logs are in the following:
+
+.. literalinclude:: ./code/generate-int-8-scale-table-for-conv-emformer.txt
+
+It generates the following two files:
+
+.. code-block:: bash
+
+  $ ls -lh encoder-scale-table.txt joiner-scale-table.txt
+  -rw-r--r-- 1 kuangfangjun root 955K Jan 11 17:28 encoder-scale-table.txt
+  -rw-r--r-- 1 kuangfangjun root  18K Jan 11 17:28 joiner-scale-table.txt
+
+.. caution::
+
+  Definitely, you need more calibration data to compute the scale table.
+
+Finally, let us use the scale table to quantize our models into ``int8``.
+
+.. code-block:: bash
+
+  ncnn2int8
+
+  usage: ncnn2int8 [inparam] [inbin] [outparam] [outbin] [calibration table]
+
+First, we quantize the encoder model:
+
+.. code-block:: bash
+
+  cd egs/librispeech/ASR
+  cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+  ncnn2int8 \
+    ./encoder_jit_trace-pnnx.ncnn.param \
+    ./encoder_jit_trace-pnnx.ncnn.bin \
+    ./encoder_jit_trace-pnnx.ncnn.int8.param \
+    ./encoder_jit_trace-pnnx.ncnn.int8.bin \
+    ./encoder-scale-table.txt
+
+Next, we quantize the joiner model:
+
+.. code-block:: bash
+
+  ncnn2int8 \
+    ./joiner_jit_trace-pnnx.ncnn.param \
+    ./joiner_jit_trace-pnnx.ncnn.bin \
+    ./joiner_jit_trace-pnnx.ncnn.int8.param \
+    ./joiner_jit_trace-pnnx.ncnn.int8.bin \
+    ./joiner-scale-table.txt
+
+The above two commands generate the following 4 files:
+
+.. code-block:: bash
+
+  -rw-r--r-- 1 kuangfangjun root  99M Jan 11 17:34 encoder_jit_trace-pnnx.ncnn.int8.bin
+  -rw-r--r-- 1 kuangfangjun root  78K Jan 11 17:34 encoder_jit_trace-pnnx.ncnn.int8.param
+  -rw-r--r-- 1 kuangfangjun root 774K Jan 11 17:35 joiner_jit_trace-pnnx.ncnn.int8.bin
+  -rw-r--r-- 1 kuangfangjun root  496 Jan 11 17:35 joiner_jit_trace-pnnx.ncnn.int8.param
+
+Congratulations! You have successfully quantized your model from ``float32`` to ``int8``.
+
+.. caution::
+
+  ``ncnn.int8.param`` and ``ncnn.int8.bin`` must be used in pairs.
+
+  You can replace ``ncnn.param`` and ``ncnn.bin`` with ``ncnn.int8.param``
+  and ``ncnn.int8.bin`` in `sherpa-ncnn`_ if you like.
+
+  For instance, to use only the ``int8`` encoder in ``sherpa-ncnn``, you can
+  replace the following invocation:
+
+    .. code-block::
+
+      cd egs/librispeech/ASR
+      cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+      sherpa-ncnn \
+        ../data/lang_bpe_500/tokens.txt \
+        ./encoder_jit_trace-pnnx.ncnn.param \
+        ./encoder_jit_trace-pnnx.ncnn.bin \
+        ./decoder_jit_trace-pnnx.ncnn.param \
+        ./decoder_jit_trace-pnnx.ncnn.bin \
+        ./joiner_jit_trace-pnnx.ncnn.param \
+        ./joiner_jit_trace-pnnx.ncnn.bin \
+        ../test_wavs/1089-134686-0001.wav
+
+  with
+
+    .. code-block::
+
+      cd egs/librispeech/ASR
+      cd icefall-asr-librispeech-conv-emformer-transducer-stateless2-2022-07-05/exp/
+
+      sherpa-ncnn \
+        ../data/lang_bpe_500/tokens.txt \
+        ./encoder_jit_trace-pnnx.ncnn.int8.param \
+        ./encoder_jit_trace-pnnx.ncnn.int8.bin \
+        ./decoder_jit_trace-pnnx.ncnn.param \
+        ./decoder_jit_trace-pnnx.ncnn.bin \
+        ./joiner_jit_trace-pnnx.ncnn.param \
+        ./joiner_jit_trace-pnnx.ncnn.bin \
+        ../test_wavs/1089-134686-0001.wav
+
+
+The following table compares again the file sizes:
+
+
++----------------------------------------+------------+
+| File name                              | File size  |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.pt              | 283 MB     |
++----------------------------------------+------------+
+| decoder_jit_trace-pnnx.pt              | 1010 KB    |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.pt               | 3.0 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin (fp16) | 142 MB     |
++----------------------------------------+------------+
+| decoder_jit_trace-pnnx.ncnn.bin (fp16) | 503 KB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin  (fp16) | 1.5 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.bin (fp32) | 283 MB     |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.bin  (fp32) | 3.0 MB     |
++----------------------------------------+------------+
+| encoder_jit_trace-pnnx.ncnn.int8.bin   | 99 MB      |
++----------------------------------------+------------+
+| joiner_jit_trace-pnnx.ncnn.int8.bin    | 774 KB     |
++----------------------------------------+------------+
+
+You can see that the file sizes of the model after ``int8`` quantization
+are much smaller.
+
+.. hint::
+
+    Currently, only linear layers and convolutional layers are quantized
+    with ``int8``, so you don't see an exact ``4x`` reduction in file sizes.
+
+.. note::
+
+  You need to test the recognition accuracy after ``int8`` quantization.
+
+You can find the speed comparison at `<https://github.com/k2-fsa/sherpa-ncnn/issues/44>`_.
+
+
+That's it! Have fun with `sherpa-ncnn`_!

docs/source/model-export/export-with-torch-jit-script.rst

@@ -1,7 +1,7 @@
 .. _export-model-with-torch-jit-script:
 
 Export model with torch.jit.script()
-===================================
+====================================
 
 In this section, we describe how to export a model via
 ``torch.jit.script()``.

docs/source/recipes/Non-streaming-ASR/aishell/conformer_ctc.rst

@@ -703,7 +703,7 @@ It will show you the following message:
 
 
 HLG decoding
-^^^^^^^^^^^^
+~~~~~~~~~~~~
 
 .. code-block:: bash
 

docs/source/recipes/Non-streaming-ASR/aishell/index.rst

@@ -19,4 +19,3 @@ It can be downloaded from `<https://www.openslr.org/33/>`_
    tdnn_lstm_ctc
    conformer_ctc
    stateless_transducer
-

docs/source/recipes/Non-streaming-ASR/aishell/tdnn_lstm_ctc.rst

@@ -498,7 +498,7 @@ We do provide a colab notebook for this recipe showing how to use a pre-trained
 |aishell asr conformer ctc colab notebook|
 
 .. |aishell asr conformer ctc colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
-   :target: https://colab.research.google.com/drive/1qULaGvXq7PCu_P61oubfz9b53JzY4H3z
+   :target: https://colab.research.google.com/drive/1jbyzYq3ytm6j2nlEt-diQm-6QVWyDDEa?usp=sharing
 
 **Congratulations!** You have finished the aishell ASR recipe with
 TDNN-LSTM CTC models in ``icefall``.

docs/source/recipes/Non-streaming-ASR/index.rst

@@ -0,0 +1,10 @@
+Non Streaming ASR
+=================
+
+.. toctree::
+   :maxdepth: 2
+
+   aishell/index
+   librispeech/index
+   timit/index
+   yesno/index

docs/source/recipes/Non-streaming-ASR/librispeech/conformer_ctc.rst

@@ -888,7 +888,7 @@ It will show you the following message:
 
 
 CTC decoding
-^^^^^^^^^^^^
+~~~~~~~~~~~~
 
 .. code-block:: bash
 
@@ -926,7 +926,7 @@ Its output is:
   YET THESE THOUGHTS AFFECTED HESTER PRYNNE LESS WITH HOPE THAN APPREHENSION
 
 HLG decoding
-^^^^^^^^^^^^
+~~~~~~~~~~~~
 
 .. code-block:: bash
 
@@ -966,7 +966,7 @@ The output is:
 
 
 HLG decoding + n-gram LM rescoring
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. code-block:: bash
 
@@ -1012,7 +1012,7 @@ The output is:
 
 
 HLG decoding + n-gram LM rescoring + attention decoder rescoring
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. code-block:: bash
 

docs/source/recipes/Non-streaming-ASR/librispeech/distillation.rst

@@ -0,0 +1,223 @@
+Distillation with HuBERT
+========================
+
+This tutorial shows you how to perform knowledge distillation in `icefall`_
+with the `LibriSpeech`_ dataset. The distillation method
+used here is called "Multi Vector Quantization Knowledge Distillation" (MVQ-KD).
+Please have a look at our paper `Predicting Multi-Codebook Vector Quantization Indexes for Knowledge Distillation <https://arxiv.org/abs/2211.00508>`_
+for more details about MVQ-KD.
+
+.. note::
+
+    This tutorial is based on recipe
+    `pruned_transducer_stateless4 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless4>`_.
+    Currently, we only implement MVQ-KD in this recipe. However, MVQ-KD is theoretically applicable to all recipes
+    with only minor changes needed. Feel free to try out MVQ-KD in different recipes. If you
+    encounter any problems, please open an issue here `icefall <https://github.com/k2-fsa/icefall/issues>`_.
+
+.. note::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for `icefall`_.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+Data preparation
+----------------
+
+We first prepare necessary training data for `LibriSpeech`_.
+This is the same as in :ref:`non_streaming_librispeech_pruned_transducer_stateless`.
+
+.. hint::
+
+   The data preparation is the same as other recipes on LibriSpeech dataset,
+   if you have finished this step, you can skip to :ref:`codebook_index_preparation` directly.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+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
+
+.. HINT::
+
+  If you have pre-downloaded the `LibriSpeech`_
+  dataset and the `musan`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+
+.. _codebook_index_preparation:
+
+Codebook index preparation
+--------------------------
+
+Here, we prepare necessary data for MVQ-KD. This requires the generation
+of codebook indexes (please read our `paper <https://arxiv.org/abs/2211.00508>`_.
+if you are interested in details). In this tutorial, we use the pre-computed
+codebook indexes for convenience. The only thing you need to do is to
+run `./distillation_with_hubert.sh <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/distillation_with_hubert.sh>`_.
+
+.. note::
+
+  There are 5 stages in total, the first and second stage will be automatically skipped
+  when choosing to downloaded codebook indexes prepared by `icefall`_.
+  Of course, you can extract and compute the codebook indexes by yourself. This
+  will require you downloading a HuBERT-XL model and it can take a while for
+  the extraction of codebook indexes.
+
+
+As usual, you can control the stages you want to run by specifying the following
+two options:
+
+  - ``--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
+
+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.
+
+- ``--full_libri`` If True, use full 960h data. Otherwise only ``train-clean-100`` will be used
+- ``--use_extracted_codebook`` If True, the first two stages will be skipped and the codebook
+  indexes uploaded by us will be downloaded.
+
+Since we are using the pre-computed codebook indexes, we set
+``use_extracted_codebook=True``. If you want to do full `LibriSpeech`_
+experiments, please set ``full_libri=True``.
+
+The following command downloads the pre-computed codebook indexes
+and prepares MVQ-augmented training manifests.
+
+.. code-block:: bash
+
+  $ ./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.
+
+.. figure:: ./images/distillation_codebook.png
+  :width: 800
+  :alt: Downloading codebook indexes and preparing training manifest.
+  :align: center
+
+  Downloading codebook indexes and preparing training manifest.
+
+.. hint::
+
+  The codebook indexes we prepared for you in this tutorial
+  are extracted from the 36-th layer of a fine-tuned HuBERT-XL model
+  with 8 codebooks. If you want to try other configurations, please
+  set ``use_extracted_codebook=False`` and set ``embedding_layer`` and
+  ``num_codebooks`` by yourself.
+
+Now, you should see the following files under the directory ``./data/vq_fbank_layer36_cb8``.
+
+.. figure:: ./images/distillation_directory.png
+  :width: 800
+  :alt: MVQ-augmented training manifests
+  :align: center
+
+  MVQ-augmented training manifests.
+
+Whola! You are ready to perform knowledge distillation training now!
+
+Training
+--------
+
+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
+
+Here is the code snippet for training:
+
+.. code-block:: bash
+
+  WORLD_SIZE=$(echo ${CUDA_VISIBLE_DEVICES} | awk '{n=split($1, _, ","); print n}')
+
+  ./pruned_transducer_stateless6/train.py \
+    --manifest-dir ./data/vq_fbank_layer36_cb8 \
+    --master-port 12359 \
+    --full-libri $full_libri \
+    --spec-aug-time-warp-factor -1 \
+    --max-duration 300 \
+    --world-size ${WORLD_SIZE} \
+    --num-epochs 30 \
+    --exp-dir $exp_dir \
+    --enable-distillation True \
+    --codebook-loss-scale 0.01
+
+There are a few training arguments in the following
+training commands that should be paid attention to.
+
+  - ``--enable-distillation`` If True, knowledge distillation training is enabled.
+  - ``--codebook-loss-scale`` The scale of the knowledge distillation loss.
+  - ``--manifest-dir`` The path to the MVQ-augmented manifest.
+
+
+Decoding
+--------
+
+After training finished, you can test the performance on using
+the following command.
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES=0
+  ./pruned_transducer_stateless6/train.py \
+    --decoding-method "modified_beam_search" \
+    --epoch 30 \
+    --avg 10 \
+    --max-duration 200 \
+    --exp-dir $exp_dir \
+    --enable-distillation True
+
+You should get similar results as `here <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS-100hours.md#distillation-with-hubert>`_.
+
+That's all! Feel free to experiment with your own setups and report your results.
+If you encounter any problems during training, please open up an issue `here <https://github.com/k2-fsa/icefall/issues>`_.

docs/source/recipes/Non-streaming-ASR/librispeech/index.rst

@@ -0,0 +1,12 @@
+LibriSpeech
+===========
+
+.. toctree::
+   :maxdepth: 1
+
+   tdnn_lstm_ctc
+   conformer_ctc
+   pruned_transducer_stateless
+   zipformer_mmi
+   zipformer_ctc_blankskip
+   distillation

docs/source/recipes/Non-streaming-ASR/librispeech/pruned_transducer_stateless.rst

@@ -0,0 +1,548 @@
+.. _non_streaming_librispeech_pruned_transducer_stateless:
+
+Pruned transducer statelessX
+============================
+
+This tutorial shows you how to run a conformer transducer model
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+.. Note::
+
+   The tutorial is suitable for `pruned_transducer_stateless <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless>`_,
+   `pruned_transducer_stateless2 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless2>`_,
+   `pruned_transducer_stateless4 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless4>`_,
+   `pruned_transducer_stateless5 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless5>`_,
+   We will take pruned_transducer_stateless4 as an example in this tutorial.
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+.. hint::
+
+   Please scroll down to the bottom of this page to find download links
+   for pretrained models if you don't want to train a model from scratch.
+
+
+We use pruned RNN-T to compute the loss.
+
+.. note::
+
+   You can find the paper about pruned RNN-T at the following address:
+
+   `<https://arxiv.org/abs/2206.13236>`_
+
+The transducer model consists of 3 parts:
+
+  - Encoder, a.k.a, the transcription network. We use a Conformer model (the reworked version by Daniel Povey)
+  - Decoder, a.k.a, the prediction network. We use a stateless model consisting of
+    ``nn.Embedding`` and ``nn.Conv1d``
+  - Joiner, a.k.a, the joint network.
+
+.. caution::
+
+   Contrary to the conventional RNN-T models, we use a stateless decoder.
+   That is, it has no recurrent connections.
+
+
+Data preparation
+----------------
+
+.. hint::
+
+   The data preparation is the same as other recipes on LibriSpeech dataset,
+   if you have finished this step, you can skip to ``Training`` directly.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. HINT::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+
+Training
+--------
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless4/train.py --help
+
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--exp-dir``
+
+    The directory to save checkpoints, training logs and tensorboard.
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./pruned_transducer_stateless4/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-1.pt``, ``epoch-2.pt``, ..., ``epoch-30.pt``
+    in the folder ``./pruned_transducer_stateless4/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./pruned_transducer_stateless4/train.py --start-epoch 10`` loads the
+    checkpoint ``./pruned_transducer_stateless4/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./pruned_transducer_stateless4/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./pruned_transducer_stateless4/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./pruned_transducer_stateless4/train.py --world-size 1
+
+    .. caution::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+  - ``--use-fp16``
+
+    If it is True, the model will train with half precision, from our experiment
+    results, by using half precision you can train with two times larger ``--max-duration``
+    so as to get almost 2X speed up.
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., number of encoder layers,
+encoder dimension, decoder dimension, number of warmup steps etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`pruned_transducer_stateless4/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless4/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./pruned_transducer_stateless4/train.py`` directly.
+
+
+.. NOTE::
+
+  The options for `pruned_transducer_stateless5 <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless5/train.py>`_ are a little different from
+  other recipes. It allows you to configure ``--num-encoder-layers``, ``--dim-feedforward``, ``--nhead``, ``--encoder-dim``, ``--decoder-dim``, ``--joiner-dim`` from commandline, so that you can train models with different size with pruned_transducer_stateless5.
+
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``--exp-dir`` (e.g. ``pruned_transducer_stateless4/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-1.pt``, ``epoch-2.pt``, ...
+
+    These are checkpoint files saved at the end of each epoch, containing model
+    ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless4/train.py --start-epoch 11
+
+  - ``checkpoint-436000.pt``, ``checkpoint-438000.pt``, ...
+
+    These are checkpoint files saved every ``--save-every-n`` batches,
+    containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``checkpoint-436000``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless4/train.py --start-batch 436000
+
+  - ``tensorboard/``
+
+    This folder contains tensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd pruned_transducer_stateless4/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "pruned transducer training for LibriSpeech with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/QOGSPBgsR8KzcRMmie9JGw/
+
+        [2022-11-20T15:50:50] Started scanning logdir.
+        Uploading 4468 scalars...
+        [2022-11-20T15:53:02] Total uploaded: 210171 scalars, 0 tensors, 0 binary objects
+        Listening for new data in logdir...
+
+    Note there is a URL in the above output. Click it and you will see
+    the following screenshot:
+
+      .. figure:: images/librispeech-pruned-transducer-tensorboard-log.jpg
+         :width: 600
+         :alt: TensorBoard screenshot
+         :align: center
+         :target: https://tensorboard.dev/experiment/QOGSPBgsR8KzcRMmie9JGw/
+
+         TensorBoard screenshot.
+
+  .. hint::
+
+    If you don't have access to google, you can use the following command
+    to view the tensorboard log locally:
+
+      .. code-block:: bash
+
+        cd pruned_transducer_stateless4/exp/tensorboard
+        tensorboard --logdir . --port 6008
+
+    It will print the following message:
+
+      .. code-block::
+
+        Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
+        TensorBoard 2.8.0 at http://localhost:6008/ (Press CTRL+C to quit)
+
+    Now start your browser and go to `<http://localhost:6008>`_ to view the tensorboard
+    logs.
+
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage example
+~~~~~~~~~~~~~
+
+You can use the following command to start the training using 6 GPUs:
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5"
+  ./pruned_transducer_stateless4/train.py \
+     --world-size 6 \
+     --num-epochs 30 \
+     --start-epoch 1 \
+     --exp-dir pruned_transducer_stateless4/exp \
+     --full-libri 1 \
+     --max-duration 300
+
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. hint::
+
+   There are two kinds of checkpoints:
+
+    - (1) ``epoch-1.pt``, ``epoch-2.pt``, ..., which are saved at the end
+      of each epoch. You can pass ``--epoch`` to
+      ``pruned_transducer_stateless4/decode.py`` to use them.
+
+    - (2) ``checkpoints-436000.pt``, ``epoch-438000.pt``, ..., which are saved
+      every ``--save-every-n`` batches. You can pass ``--iter`` to
+      ``pruned_transducer_stateless4/decode.py`` to use them.
+
+    We suggest that you try both types of checkpoints and choose the one
+    that produces the lowest WERs.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless4/decode.py --help
+
+shows the options for decoding.
+
+The following shows two examples (for two types of checkpoints):
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for epoch in 25 20; do
+      for avg in 7 5 3 1; do
+        ./pruned_transducer_stateless4/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --exp-dir pruned_transducer_stateless4/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for iter in 474000; do
+      for avg in 8 10 12 14 16 18; do
+        ./pruned_transducer_stateless4/decode.py \
+          --iter $iter \
+          --avg $avg \
+          --exp-dir pruned_transducer_stateless4/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. Note::
+
+  Supporting decoding methods are as follows:
+
+    - ``greedy_search`` : It takes the symbol with largest posterior probability
+      of each frame as the decoding result.
+
+    - ``beam_search`` :  It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf and
+      `espnet/nets/beam_search_transducer.py <https://github.com/espnet/espnet/blob/master/espnet/nets/beam_search_transducer.py#L247>`_
+      is used as a reference. Basicly, it keeps topk states for each frame, and expands the kept states with their own contexts to
+      next frame.
+
+    - ``modified_beam_search`` : It implements the same algorithm as ``beam_search`` above, but it
+      runs in batch mode with ``--max-sym-per-frame=1`` being hardcoded.
+
+    - ``fast_beam_search`` : It implements graph composition between the output ``log_probs`` and
+      given ``FSAs``. It is hard to describe the details in several lines of texts, you can read
+      our paper in https://arxiv.org/pdf/2211.00484.pdf or our `rnnt decode code in k2 <https://github.com/k2-fsa/k2/blob/master/k2/csrc/rnnt_decode.h>`_. ``fast_beam_search`` can decode with ``FSAs`` on GPU efficiently.
+
+    - ``fast_beam_search_LG`` : The same as ``fast_beam_search`` above, ``fast_beam_search`` uses
+      an trivial graph that has only one state, while ``fast_beam_search_LG`` uses an LG graph
+      (with N-gram LM).
+
+    - ``fast_beam_search_nbest`` : It produces the decoding results as follows:
+
+      - (1) Use ``fast_beam_search`` to get a lattice
+      - (2) Select ``num_paths`` paths from the lattice using ``k2.random_paths()``
+      - (3) Unique the selected paths
+      - (4) Intersect the selected paths with the lattice and compute the
+            shortest path from the intersection result
+      - (5) The path with the largest score is used as the decoding output.
+
+    - ``fast_beam_search_nbest_LG`` : It implements same logic as ``fast_beam_search_nbest``, the
+      only difference is that it uses ``fast_beam_search_LG`` to generate the lattice.
+
+
+Export Model
+------------
+
+`pruned_transducer_stateless4/export.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless4/export.py>`_ supports exporting checkpoints from ``pruned_transducer_stateless4/exp`` in the following ways.
+
+Export ``model.state_dict()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Checkpoints saved by ``pruned_transducer_stateless4/train.py`` also include
+``optimizer.state_dict()``. It is useful for resuming training. But after training,
+we are interested only in ``model.state_dict()``. You can use the following
+command to extract ``model.state_dict()``.
+
+.. code-block:: bash
+
+  # Assume that --epoch 25 --avg 3 produces the smallest WER
+  # (You can get such information after running ./pruned_transducer_stateless4/decode.py)
+
+  epoch=25
+  avg=3
+
+  ./pruned_transducer_stateless4/export.py \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch $epoch \
+    --avg  $avg
+
+It will generate a file ``./pruned_transducer_stateless4/exp/pretrained.pt``.
+
+.. hint::
+
+   To use the generated ``pretrained.pt`` for ``pruned_transducer_stateless4/decode.py``,
+   you can run:
+
+   .. code-block:: bash
+
+      cd pruned_transducer_stateless4/exp
+      ln -s pretrained.pt epoch-999.pt
+
+   And then pass ``--epoch 999 --avg 1 --use-averaged-model 0`` to
+   ``./pruned_transducer_stateless4/decode.py``.
+
+To use the exported model with ``./pruned_transducer_stateless4/pretrained.py``, you
+can run:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless4/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless4/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+
+Export model using ``torch.jit.script()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless4/export.py \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 25 \
+    --avg 3 \
+    --jit 1
+
+It will generate a file ``cpu_jit.pt`` in the given ``exp_dir``. You can later
+load it by ``torch.jit.load("cpu_jit.pt")``.
+
+Note ``cpu`` in the name ``cpu_jit.pt`` means the parameters when loaded into Python
+are on CPU. You can use ``to("cuda")`` to move them to a CUDA device.
+
+.. NOTE::
+
+   You will need this ``cpu_jit.pt`` when deploying with Sherpa framework.
+
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following links:
+
+  - `pruned_transducer_stateless <https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless-2022-03-12>`_
+
+  - `pruned_transducer_stateless2 <https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless2-2022-04-29>`_
+
+  - `pruned_transducer_stateless4 <https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless4-2022-06-03>`_
+
+  - `pruned_transducer_stateless5 <https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless5-2022-07-07>`_
+
+  See `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  for the details of the above pretrained models
+
+
+Deploy with Sherpa
+------------------
+
+Please see `<https://k2-fsa.github.io/sherpa/python/offline_asr/conformer/librispeech.html#>`_
+for how to deploy the models in ``sherpa``.

docs/source/recipes/Non-streaming-ASR/librispeech/tdnn_lstm_ctc.rst

@@ -398,7 +398,7 @@ We provide a colab notebook for decoding with pre-trained model.
 |librispeech tdnn_lstm_ctc colab notebook|
 
 .. |librispeech tdnn_lstm_ctc colab notebook| image:: https://colab.research.google.com/assets/colab-badge.svg
-   :target: https://colab.research.google.com/drive/1kNmDXNMwREi0rZGAOIAOJo93REBuOTcd
+   :target: https://colab.research.google.com/drive/1-iSfQMp2So-We_Uu49N4AAcMInB72u9z?usp=sharing
 
 
 **Congratulations!** You have finished the TDNN-LSTM-CTC recipe on librispeech in ``icefall``.

docs/source/recipes/Non-streaming-ASR/librispeech/zipformer_ctc_blankskip.rst

@@ -0,0 +1,454 @@
+Zipformer CTC Blank Skip
+========================
+
+.. hint::
+
+   Please scroll down to the bottom of this page to find download links
+   for pretrained models if you don't want to train a model from scratch.
+
+
+This tutorial shows you how to train a Zipformer model based on the guidance from 
+a co-trained CTC model using `blank skip method <https://arxiv.org/pdf/2210.16481.pdf>`_
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+.. note::
+
+    We use both CTC and RNN-T loss to train. During the forward pass, the encoder output
+    is first used to calculate the CTC posterior probability; then for each output frame,
+    if its blank posterior is bigger than some threshold, it will be simply discarded
+    from the encoder output. To prevent information loss, we also put a convolution module
+    similar to the one used in conformer (referred to as “LConv”) before the frame reduction.
+
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+.. note::
+
+   We encourage you to read ``./prepare.sh``.
+
+The data preparation contains several stages. You can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. hint::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. note::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+Training
+--------
+
+For stability, it doesn`t use blank skip method until model warm-up.
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless7_ctc_bs/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./pruned_transducer_stateless7_ctc_bs/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-1.pt``, ``epoch-2.pt``, ..., ``epoch-30.pt``
+    in the folder ``./pruned_transducer_stateless7_ctc_bs/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./pruned_transducer_stateless7_ctc_bs/train.py --start-epoch 10`` loads the
+    checkpoint ``./pruned_transducer_stateless7_ctc_bs/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./pruned_transducer_stateless7_ctc_bs/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./pruned_transducer_stateless7_ctc_bs/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./pruned_transducer_stateless7_ctc_bs/train.py --world-size 1
+
+    .. caution::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., weight decay,
+number of warmup steps, results dir, etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`pruned_transducer_stateless7_ctc_bs/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./pruned_transducer_stateless7_ctc_bs/train.py`` directly.
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``pruned_transducer_stateless7_ctc_bs/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-1.pt``, ``epoch-2.pt``, ...
+
+    These are checkpoint files saved at the end of each epoch, containing model
+    ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless7_ctc_bs/train.py --start-epoch 11
+
+  - ``checkpoint-436000.pt``, ``checkpoint-438000.pt``, ...
+
+    These are checkpoint files saved every ``--save-every-n`` batches,
+    containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``checkpoint-436000``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless7_ctc_bs/train.py --start-batch 436000
+
+  - ``tensorboard/``
+
+    This folder contains tensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd pruned_transducer_stateless7_ctc_bs/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "Zipformer-CTC co-training using blank skip for LibriSpeech with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/xyOZUKpEQm62HBIlUD4uPA/
+
+    Note there is a URL in the above output. Click it and you will see
+    tensorboard.
+
+  .. hint::
+
+    If you don't have access to google, you can use the following command
+    to view the tensorboard log locally:
+
+      .. code-block:: bash
+
+        cd pruned_transducer_stateless7_ctc_bs/exp/tensorboard
+        tensorboard --logdir . --port 6008
+
+    It will print the following message:
+
+      .. code-block::
+
+        Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
+        TensorBoard 2.8.0 at http://localhost:6008/ (Press CTRL+C to quit)
+
+    Now start your browser and go to `<http://localhost:6008>`_ to view the tensorboard
+    logs.
+
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage example
+~~~~~~~~~~~~~
+
+You can use the following command to start the training using 4 GPUs:
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./pruned_transducer_stateless7_ctc_bs/train.py \
+    --world-size 4 \
+    --num-epochs 30 \
+    --start-epoch 1 \
+    --full-libri 1 \
+    --exp-dir pruned_transducer_stateless7_ctc_bs/exp \
+    --max-duration 600 \
+    --use-fp16 1
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. hint::
+
+   There are two kinds of checkpoints:
+
+    - (1) ``epoch-1.pt``, ``epoch-2.pt``, ..., which are saved at the end
+      of each epoch. You can pass ``--epoch`` to
+      ``pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py`` to use them.
+
+    - (2) ``checkpoints-436000.pt``, ``epoch-438000.pt``, ..., which are saved
+      every ``--save-every-n`` batches. You can pass ``--iter`` to
+      ``pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py`` to use them.
+
+    We suggest that you try both types of checkpoints and choose the one
+    that produces the lowest WERs.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py --help
+
+shows the options for decoding.
+
+The following shows the example using ``epoch-*.pt``:
+
+.. code-block:: bash
+
+    for m in greedy_search fast_beam_search modified_beam_search; do
+        ./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py \
+            --epoch 30 \
+            --avg 13 \
+            --exp-dir pruned_transducer_stateless7_ctc_bs/exp \
+            --max-duration 600 \
+            --decoding-method $m
+    done
+
+To test CTC branch, you can use the following command:
+
+.. code-block:: bash
+
+    for m in ctc-decoding 1best; do
+        ./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py \
+            --epoch 30 \
+            --avg 13 \
+            --exp-dir pruned_transducer_stateless7_ctc_bs/exp \
+            --max-duration 600 \
+            --decoding-method $m
+    done
+
+Export models
+-------------
+
+`pruned_transducer_stateless7_ctc_bs/export.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/export.py>`_ supports exporting checkpoints from ``pruned_transducer_stateless7_ctc_bs/exp`` in the following ways.
+
+Export ``model.state_dict()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Checkpoints saved by ``pruned_transducer_stateless7_ctc_bs/train.py`` also include
+``optimizer.state_dict()``. It is useful for resuming training. But after training,
+we are interested only in ``model.state_dict()``. You can use the following
+command to extract ``model.state_dict()``.
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_ctc_bs/export.py \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 13 \
+    --jit 0
+
+It will generate a file ``./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt``.
+
+.. hint::
+
+   To use the generated ``pretrained.pt`` for ``pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py``,
+   you can run:
+
+   .. code-block:: bash
+
+      cd pruned_transducer_stateless7_ctc_bs/exp
+      ln -s pretrained epoch-9999.pt
+
+   And then pass ``--epoch 9999 --avg 1 --use-averaged-model 0`` to
+   ``./pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py``.
+
+To use the exported model with ``./pruned_transducer_stateless7_ctc_bs/pretrained.py``, you
+can run:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_ctc_bs/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+To test CTC branch using the exported model with ``./pruned_transducer_stateless7_ctc_bs/pretrained_ctc.py``:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py \
+    --checkpoint ./pruned_transducer_stateless7_ctc_bs/exp/pretrained.pt \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --method ctc-decoding \
+    --sample-rate 16000 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+Export model using ``torch.jit.script()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_ctc_bs/export.py \
+    --exp-dir ./pruned_transducer_stateless7_ctc_bs/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 13 \
+    --jit 1
+
+It will generate a file ``cpu_jit.pt`` in the given ``exp_dir``. You can later
+load it by ``torch.jit.load("cpu_jit.pt")``.
+
+Note ``cpu`` in the name ``cpu_jit.pt`` means the parameters when loaded into Python
+are on CPU. You can use ``to("cuda")`` to move them to a CUDA device.
+
+To use the generated files with ``./pruned_transducer_stateless7_ctc_bs/jit_pretrained.py``:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_ctc_bs/jit_pretrained.py \
+    --nn-model-filename ./pruned_transducer_stateless7_ctc_bs/exp/cpu_jit.pt \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+To test CTC branch using the generated files with ``./pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py``:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_ctc_bs/jit_pretrained_ctc.py \
+    --model-filename ./pruned_transducer_stateless7_ctc_bs/exp/cpu_jit.pt \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --method ctc-decoding \
+    --sample-rate 16000 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following links:
+
+  - trained on LibriSpeech 100h: `<https://huggingface.co/yfyeung/icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2022-12-14>`_
+  - trained on LibriSpeech 960h: `<https://huggingface.co/yfyeung/icefall-asr-librispeech-pruned_transducer_stateless7_ctc_bs-2023-01-29>`_
+
+  See `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  for the details of the above pretrained models

docs/source/recipes/Non-streaming-ASR/librispeech/zipformer_mmi.rst

@@ -0,0 +1,422 @@
+Zipformer MMI
+===============
+
+.. hint::
+
+   Please scroll down to the bottom of this page to find download links
+   for pretrained models if you don't want to train a model from scratch.
+
+
+This tutorial shows you how to train an Zipformer MMI model
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+We use LF-MMI to compute the loss.
+
+.. note::
+
+   You can find the document about LF-MMI training at the following address:
+
+   `<https://github.com/k2-fsa/next-gen-kaldi-wechat/blob/master/pdf/LF-MMI-training-and-decoding-in-k2-Part-I.pdf>`_
+
+
+Data preparation
+----------------
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+.. note::
+
+   We encourage you to read ``./prepare.sh``.
+
+The data preparation contains several stages. You can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. hint::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. note::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+Training
+--------
+
+For stability, it uses CTC loss for model warm-up and then switches to MMI loss.
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./zipformer_mmi/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./zipformer_mmi/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-1.pt``, ``epoch-2.pt``, ..., ``epoch-30.pt``
+    in the folder ``./zipformer_mmi/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./zipformer_mmi/train.py --start-epoch 10`` loads the
+    checkpoint ``./zipformer_mmi/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./zipformer_mmi/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./zipformer_mmi/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./zipformer_mmi/train.py --world-size 1
+
+    .. caution::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., weight decay,
+number of warmup steps, results dir, etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`zipformer_mmi/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/zipformer_mmi/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./zipformer_mmi/train.py`` directly.
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``zipformer_mmi/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-1.pt``, ``epoch-2.pt``, ...
+
+    These are checkpoint files saved at the end of each epoch, containing model
+    ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./zipformer_mmi/train.py --start-epoch 11
+
+  - ``checkpoint-436000.pt``, ``checkpoint-438000.pt``, ...
+
+    These are checkpoint files saved every ``--save-every-n`` batches,
+    containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``checkpoint-436000``, you can use:
+
+      .. code-block:: bash
+
+        $ ./zipformer_mmi/train.py --start-batch 436000
+
+  - ``tensorboard/``
+
+    This folder contains tensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd zipformer_mmi/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "Zipformer MMI training for LibriSpeech with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/xyOZUKpEQm62HBIlUD4uPA/
+
+    Note there is a URL in the above output. Click it and you will see
+    tensorboard.
+
+  .. hint::
+
+    If you don't have access to google, you can use the following command
+    to view the tensorboard log locally:
+
+      .. code-block:: bash
+
+        cd zipformer_mmi/exp/tensorboard
+        tensorboard --logdir . --port 6008
+
+    It will print the following message:
+
+      .. code-block::
+
+        Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
+        TensorBoard 2.8.0 at http://localhost:6008/ (Press CTRL+C to quit)
+
+    Now start your browser and go to `<http://localhost:6008>`_ to view the tensorboard
+    logs.
+
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage example
+~~~~~~~~~~~~~
+
+You can use the following command to start the training using 4 GPUs:
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./zipformer_mmi/train.py \
+    --world-size 4 \
+    --num-epochs 30 \
+    --start-epoch 1 \
+    --full-libri 1 \
+    --exp-dir zipformer_mmi/exp \
+    --max-duration 500 \
+    --use-fp16 1 \
+    --num-workers 2
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. hint::
+
+   There are two kinds of checkpoints:
+
+    - (1) ``epoch-1.pt``, ``epoch-2.pt``, ..., which are saved at the end
+      of each epoch. You can pass ``--epoch`` to
+      ``zipformer_mmi/decode.py`` to use them.
+
+    - (2) ``checkpoints-436000.pt``, ``epoch-438000.pt``, ..., which are saved
+      every ``--save-every-n`` batches. You can pass ``--iter`` to
+      ``zipformer_mmi/decode.py`` to use them.
+
+    We suggest that you try both types of checkpoints and choose the one
+    that produces the lowest WERs.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./zipformer_mmi/decode.py --help
+
+shows the options for decoding.
+
+The following shows the example using ``epoch-*.pt``:
+
+.. code-block:: bash
+
+  for m in nbest nbest-rescoring-LG nbest-rescoring-3-gram nbest-rescoring-4-gram; do
+    ./zipformer_mmi/decode.py \
+      --epoch 30 \
+      --avg 10 \
+      --exp-dir ./zipformer_mmi/exp/ \
+      --max-duration 100 \
+      --lang-dir data/lang_bpe_500 \
+      --nbest-scale 1.2 \
+      --hp-scale 1.0 \
+      --decoding-method $m
+  done
+
+
+Export models
+-------------
+
+`zipformer_mmi/export.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/zipformer_mmi/export.py>`_ supports exporting checkpoints from ``zipformer_mmi/exp`` in the following ways.
+
+Export ``model.state_dict()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Checkpoints saved by ``zipformer_mmi/train.py`` also include
+``optimizer.state_dict()``. It is useful for resuming training. But after training,
+we are interested only in ``model.state_dict()``. You can use the following
+command to extract ``model.state_dict()``.
+
+.. code-block:: bash
+
+  ./zipformer_mmi/export.py \
+    --exp-dir ./zipformer_mmi/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 9 \
+    --jit 0
+
+It will generate a file ``./zipformer_mmi/exp/pretrained.pt``.
+
+.. hint::
+
+   To use the generated ``pretrained.pt`` for ``zipformer_mmi/decode.py``,
+   you can run:
+
+   .. code-block:: bash
+
+      cd zipformer_mmi/exp
+      ln -s pretrained epoch-9999.pt
+
+   And then pass ``--epoch 9999 --avg 1 --use-averaged-model 0`` to
+   ``./zipformer_mmi/decode.py``.
+
+To use the exported model with ``./zipformer_mmi/pretrained.py``, you
+can run:
+
+.. code-block:: bash
+
+  ./zipformer_mmi/pretrained.py \
+    --checkpoint ./zipformer_mmi/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method 1best \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+Export model using ``torch.jit.script()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./zipformer_mmi/export.py \
+    --exp-dir ./zipformer_mmi/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 9 \
+    --jit 1
+
+It will generate a file ``cpu_jit.pt`` in the given ``exp_dir``. You can later
+load it by ``torch.jit.load("cpu_jit.pt")``.
+
+Note ``cpu`` in the name ``cpu_jit.pt`` means the parameters when loaded into Python
+are on CPU. You can use ``to("cuda")`` to move them to a CUDA device.
+
+To use the generated files with ``./zipformer_mmi/jit_pretrained.py``:
+
+.. code-block:: bash
+
+  ./zipformer_mmi/jit_pretrained.py \
+    --nn-model-filename ./zipformer_mmi/exp/cpu_jit.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method 1best \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following links:
+
+  - `<https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-mmi-2022-12-08>`_
+
+  See `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  for the details of the above pretrained models

docs/source/recipes/Non-streaming-ASR/timit/index.rst

@@ -6,4 +6,3 @@ TIMIT
 
    tdnn_ligru_ctc
    tdnn_lstm_ctc
-

docs/source/recipes/Streaming-ASR/index.rst

@@ -0,0 +1,12 @@
+Streaming ASR
+=============
+
+.. toctree::
+   :maxdepth: 1
+
+   introduction
+
+.. toctree::
+   :maxdepth: 2
+
+   librispeech/index

docs/source/recipes/Streaming-ASR/introduction.rst

@@ -0,0 +1,53 @@
+Introduction
+============
+
+This page shows you how we implement streaming **X-former transducer** models for ASR.
+
+.. HINT::
+   X-former transducer here means the encoder of the transducer model uses Multi-Head Attention,
+   like `Conformer <https://arxiv.org/pdf/2005.08100.pdf>`_, `EmFormer <https://arxiv.org/pdf/2010.10759.pdf>`_ etc.
+
+Currently we have implemented two types of streaming models, one uses Conformer as encoder, the other uses Emformer as encoder.
+
+Streaming Conformer
+-------------------
+
+The main idea of training a streaming model is to make the model see limited contexts
+in training time, we can achieve this by applying a mask to the output of self-attention.
+In icefall, we implement the streaming conformer the way just like what `WeNet <https://arxiv.org/pdf/2012.05481.pdf>`_ did.
+
+.. NOTE::
+   The conformer-transducer recipes in LibriSpeech datasets, like, `pruned_transducer_stateless <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless>`_,
+   `pruned_transducer_stateless2 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless2>`_,
+   `pruned_transducer_stateless3 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless3>`_,
+   `pruned_transducer_stateless4 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless4>`_,
+   `pruned_transducer_stateless5 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless5>`_
+   all support streaming.
+
+.. NOTE::
+   Training a streaming conformer model in ``icefall`` is almost the same as training a
+   non-streaming model, all you need to do is passing several extra arguments.
+   See :doc:`Pruned transducer statelessX <librispeech/pruned_transducer_stateless>` for more details.
+
+.. HINT::
+   If you want to modify a non-streaming conformer recipe to support both streaming and non-streaming, please refer
+   to `this pull request <https://github.com/k2-fsa/icefall/pull/454>`_.  After adding the code needed by streaming training,
+   you have to re-train it with the extra arguments metioned in the docs above to get a streaming model.
+
+
+Streaming Emformer
+------------------
+
+The Emformer model proposed `here <https://arxiv.org/pdf/2010.10759.pdf>`_ uses more
+complicated techniques. It has a memory bank component to memorize history information,
+what' more, it also introduces right context in training time by hard-copying part of
+the input features.
+
+We have three variants of Emformer models in ``icefall``.
+
+ - ``pruned_stateless_emformer_rnnt2`` using Emformer from torchaudio, see `LibriSpeech recipe <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_stateless_emformer_rnnt2>`_.
+ - ``conv_emformer_transducer_stateless`` using ConvEmformer implemented by ourself. Different from the Emformer in torchaudio,
+   ConvEmformer has a convolution in each layer and uses the mechanisms in our reworked conformer model.
+   See `LibriSpeech recipe <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/conv_emformer_transducer_stateless>`_.
+ - ``conv_emformer_transducer_stateless2`` using ConvEmformer implemented by ourself. The only difference from the above one is that
+   it uses a simplified memory bank. See `LibriSpeech recipe <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/conv_emformer_transducer_stateless2>`_.

docs/source/recipes/Streaming-ASR/librispeech/index.rst

@@ -4,6 +4,8 @@ LibriSpeech
 .. toctree::
    :maxdepth: 1
 
-   tdnn_lstm_ctc
-   conformer_ctc
+   pruned_transducer_stateless
+
    lstm_pruned_stateless_transducer
+
+   zipformer_transducer

docs/source/recipes/Streaming-ASR/librispeech/lstm_pruned_stateless_transducer.rst

@@ -515,10 +515,10 @@ To use the generated files with ``./lstm_transducer_stateless2/jit_pretrained``:
    Please see `<https://k2-fsa.github.io/sherpa/python/streaming_asr/lstm/english/server.html>`_
    for how to use the exported models in ``sherpa``.
 
-.. _export-model-for-ncnn:
+.. _export-lstm-transducer-model-for-ncnn:
 
-Export model for ncnn
-~~~~~~~~~~~~~~~~~~~~~
+Export LSTM transducer models for ncnn
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 We support exporting pretrained LSTM transducer models to
 `ncnn <https://github.com/tencent/ncnn>`_ using
@@ -531,16 +531,36 @@ First, let us install a modified version of ``ncnn``:
   git clone https://github.com/csukuangfj/ncnn
   cd ncnn
   git submodule update --recursive --init
-  python3 setup.py bdist_wheel
-  ls -lh dist/
-  pip install ./dist/*.whl
+
+  # Note: We don't use "python setup.py install" or "pip install ." here
+
+  mkdir -p build-wheel
+  cd build-wheel
+
+  cmake \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DNCNN_PYTHON=ON \
+    -DNCNN_BUILD_BENCHMARK=OFF \
+    -DNCNN_BUILD_EXAMPLES=OFF \
+    -DNCNN_BUILD_TOOLS=ON \
+    ..
+
+  make -j4
+
+  cd ..
+
+  # Note: $PWD here is /path/to/ncnn
+
+  export PYTHONPATH=$PWD/python:$PYTHONPATH
+  export PATH=$PWD/tools/pnnx/build/src:$PATH
+  export PATH=$PWD/build-wheel/tools/quantize:$PATH
 
   # now build pnnx
   cd tools/pnnx
   mkdir build
   cd build
+  cmake ..
   make -j4
-  export PATH=$PWD/src:$PATH
 
   ./src/pnnx
 
@@ -549,6 +569,9 @@ First, let us install a modified version of ``ncnn``:
    We assume that you have added the path to the binary ``pnnx`` to the
    environment variable ``PATH``.
 
+   We also assume that you have added ``build/tools/quantize`` to the environment
+   variable ``PATH`` so that you are able to use ``ncnn2int8`` later.
+
 Second, let us export the model using ``torch.jit.trace()`` that is suitable
 for ``pnnx``:
 
@@ -634,3 +657,6 @@ by visiting the following links:
 
 You can find more usages of the pretrained models in
 `<https://k2-fsa.github.io/sherpa/python/streaming_asr/lstm/index.html>`_
+
+Export ConvEmformer transducer models for ncnn
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

docs/source/recipes/Streaming-ASR/librispeech/pruned_transducer_stateless.rst

@@ -0,0 +1,735 @@
+Pruned transducer statelessX
+============================
+
+This tutorial shows you how to run a **streaming** conformer transducer model
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+.. Note::
+
+   The tutorial is suitable for `pruned_transducer_stateless <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless>`_,
+   `pruned_transducer_stateless2 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless2>`_,
+   `pruned_transducer_stateless4 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless4>`_,
+   `pruned_transducer_stateless5 <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless5>`_,
+   We will take pruned_transducer_stateless4 as an example in this tutorial.
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+.. hint::
+
+   Please scroll down to the bottom of this page to find download links
+   for pretrained models if you don't want to train a model from scratch.
+
+
+We use pruned RNN-T to compute the loss.
+
+.. note::
+
+   You can find the paper about pruned RNN-T at the following address:
+
+   `<https://arxiv.org/abs/2206.13236>`_
+
+The transducer model consists of 3 parts:
+
+  - Encoder, a.k.a, the transcription network. We use a Conformer model (the reworked version by Daniel Povey)
+  - Decoder, a.k.a, the prediction network. We use a stateless model consisting of
+    ``nn.Embedding`` and ``nn.Conv1d``
+  - Joiner, a.k.a, the joint network.
+
+.. caution::
+
+   Contrary to the conventional RNN-T models, we use a stateless decoder.
+   That is, it has no recurrent connections.
+
+
+Data preparation
+----------------
+
+.. hint::
+
+   The data preparation is the same as other recipes on LibriSpeech dataset,
+   if you have finished this step, you can skip to ``Training`` directly.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. HINT::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+
+Training
+--------
+
+.. NOTE::
+
+   We put the streaming and non-streaming model in one recipe, to train a streaming model you only
+   need to add **4** extra options comparing with training a non-streaming model. These options are
+   ``--dynamic-chunk-training``, ``--num-left-chunks``, ``--causal-convolution``, ``--short-chunk-size``.
+   You can see the configurable options below for their meanings or read https://arxiv.org/pdf/2012.05481.pdf for more details.
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless4/train.py --help
+
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--exp-dir``
+
+    The directory to save checkpoints, training logs and tensorboard.
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./pruned_transducer_stateless4/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-1.pt``, ``epoch-2.pt``, ..., ``epoch-30.pt``
+    in the folder ``./pruned_transducer_stateless4/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./pruned_transducer_stateless4/train.py --start-epoch 10`` loads the
+    checkpoint ``./pruned_transducer_stateless4/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./pruned_transducer_stateless4/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./pruned_transducer_stateless4/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./pruned_transducer_stateless4/train.py --world-size 1
+
+    .. caution::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+  - ``--use-fp16``
+
+    If it is True, the model will train with half precision, from our experiment
+    results, by using half precision you can train with two times larger ``--max-duration``
+    so as to get almost 2X speed up.
+
+  - ``--dynamic-chunk-training``
+
+    The flag that indicates whether to train a streaming model or not, it
+    **MUST** be True if you want to train a streaming model.
+
+  - ``--short-chunk-size``
+
+    When training a streaming attention model with chunk masking, the chunk size
+    would be either max sequence length of current batch or uniformly sampled from
+    (1, short_chunk_size). The default value is 25, you don't have to change it most of the time.
+
+  - ``--num-left-chunks``
+
+    It indicates how many left context (in chunks) that can be seen when calculating attention.
+    The default value is 4, you don't have to change it most of the time.
+
+
+  - ``--causal-convolution``
+
+    Whether to use causal convolution in conformer encoder layer, this requires
+    to be True when training a streaming model.
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., number of encoder layers,
+encoder dimension, decoder dimension, number of warmup steps etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`pruned_transducer_stateless4/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless4/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./pruned_transducer_stateless4/train.py`` directly.
+
+
+.. NOTE::
+
+  The options for `pruned_transducer_stateless5 <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless5/train.py>`_ are a little different from
+  other recipes. It allows you to configure ``--num-encoder-layers``, ``--dim-feedforward``, ``--nhead``, ``--encoder-dim``, ``--decoder-dim``, ``--joiner-dim`` from commandline, so that you can train models with different size with pruned_transducer_stateless5.
+
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``--exp-dir`` (e.g. ``pruned_transducer_stateless4/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-1.pt``, ``epoch-2.pt``, ...
+
+    These are checkpoint files saved at the end of each epoch, containing model
+    ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless4/train.py --start-epoch 11
+
+  - ``checkpoint-436000.pt``, ``checkpoint-438000.pt``, ...
+
+    These are checkpoint files saved every ``--save-every-n`` batches,
+    containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``checkpoint-436000``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless4/train.py --start-batch 436000
+
+  - ``tensorboard/``
+
+    This folder contains tensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd pruned_transducer_stateless4/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "pruned transducer training for LibriSpeech with icefall"
+
+    It will print something like below:
+
+      .. code-block::
+
+        TensorFlow installation not found - running with reduced feature set.
+        Upload started and will continue reading any new data as it's added to the logdir.
+
+        To stop uploading, press Ctrl-C.
+
+        New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/97VKXf80Ru61CnP2ALWZZg/
+
+        [2022-11-20T15:50:50] Started scanning logdir.
+        Uploading 4468 scalars...
+        [2022-11-20T15:53:02] Total uploaded: 210171 scalars, 0 tensors, 0 binary objects
+        Listening for new data in logdir...
+
+    Note there is a URL in the above output. Click it and you will see
+    the following screenshot:
+
+      .. figure:: images/streaming-librispeech-pruned-transducer-tensorboard-log.jpg
+         :width: 600
+         :alt: TensorBoard screenshot
+         :align: center
+         :target: https://tensorboard.dev/experiment/97VKXf80Ru61CnP2ALWZZg/
+
+         TensorBoard screenshot.
+
+  .. hint::
+
+    If you don't have access to google, you can use the following command
+    to view the tensorboard log locally:
+
+      .. code-block:: bash
+
+        cd pruned_transducer_stateless4/exp/tensorboard
+        tensorboard --logdir . --port 6008
+
+    It will print the following message:
+
+      .. code-block::
+
+        Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
+        TensorBoard 2.8.0 at http://localhost:6008/ (Press CTRL+C to quit)
+
+    Now start your browser and go to `<http://localhost:6008>`_ to view the tensorboard
+    logs.
+
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage example
+~~~~~~~~~~~~~
+
+You can use the following command to start the training using 4 GPUs:
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./pruned_transducer_stateless4/train.py \
+     --world-size 4 \
+     --dynamic-chunk-training 1 \
+     --causal-convolution 1 \
+     --num-epochs 30 \
+     --start-epoch 1 \
+     --exp-dir pruned_transducer_stateless4/exp \
+     --full-libri 1 \
+     --max-duration 300
+
+.. NOTE::
+
+   Comparing with training a non-streaming model, you only need to add two extra options,
+   ``--dynamic-chunk-training 1``  and ``--causal-convolution 1`` .
+
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. hint::
+
+   There are two kinds of checkpoints:
+
+    - (1) ``epoch-1.pt``, ``epoch-2.pt``, ..., which are saved at the end
+      of each epoch. You can pass ``--epoch`` to
+      ``pruned_transducer_stateless4/decode.py`` to use them.
+
+    - (2) ``checkpoints-436000.pt``, ``epoch-438000.pt``, ..., which are saved
+      every ``--save-every-n`` batches. You can pass ``--iter`` to
+      ``pruned_transducer_stateless4/decode.py`` to use them.
+
+    We suggest that you try both types of checkpoints and choose the one
+    that produces the lowest WERs.
+
+.. tip::
+
+    To decode a streaming model, you can use either ``simulate streaming decoding`` in ``decode.py`` or
+    ``real streaming decoding`` in ``streaming_decode.py``, the difference between ``decode.py`` and
+    ``streaming_decode.py`` is that, ``decode.py`` processes the whole acoustic frames at one time with masking (i.e. same as training),
+    but ``streaming_decode.py`` processes the acoustic frames chunk by chunk (so it can only see limited context).
+
+.. NOTE::
+
+   ``simulate streaming decoding`` in ``decode.py`` and ``real streaming decoding`` in ``streaming_decode.py`` should
+   produce almost the same results given the same ``--decode-chunk-size`` and ``--left-context``.
+
+
+Simulate streaming decoding
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless4/decode.py --help
+
+shows the options for decoding.
+The following options are important for streaming models:
+
+  ``--simulate-streaming``
+
+    If you want to decode a streaming model with ``decode.py``, you **MUST** set
+    ``--simulate-streaming`` to ``True``. ``simulate`` here means the acoustic frames
+    are not processed frame by frame (or chunk by chunk), instead, the whole sequence
+    is processed at one time with masking (the same as training).
+
+  ``--causal-convolution``
+
+    If True, the convolution module in encoder layers will be causal convolution.
+    This is **MUST** be True when decoding with a streaming model.
+
+  ``--decode-chunk-size``
+
+    For streaming models, we will calculate the chunk-wise attention, ``--decode-chunk-size``
+    indicates the chunk length (in frames after subsampling) for chunk-wise attention.
+    For ``simulate streaming decoding`` the ``decode-chunk-size`` is used to generate
+    the attention mask.
+
+  ``--left-context``
+
+    ``--left-context`` indicates how many left context frames (after subsampling) can be seen
+    for current chunk when calculating chunk-wise attention. Normally, ``left-context`` should equal
+    to ``decode-chunk-size * num-left-chunks``, where ``num-left-chunks`` is the option used
+    to train this model. For ``simulate streaming decoding`` the ``left-context`` is used to generate
+    the attention mask.
+
+
+The following shows two examples (for the two types of checkpoints):
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for epoch in 25 20; do
+      for avg in 7 5 3 1; do
+        ./pruned_transducer_stateless4/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --simulate-streaming 1 \
+          --causal-convolution 1 \
+          --decode-chunk-size 16 \
+          --left-context 64 \
+          --exp-dir pruned_transducer_stateless4/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for iter in 474000; do
+      for avg in 8 10 12 14 16 18; do
+        ./pruned_transducer_stateless4/decode.py \
+          --iter $iter \
+          --avg $avg \
+          --simulate-streaming 1 \
+          --causal-convolution 1 \
+          --decode-chunk-size 16 \
+          --left-context 64 \
+          --exp-dir pruned_transducer_stateless4/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+Real streaming decoding
+~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless4/streaming_decode.py --help
+
+shows the options for decoding.
+The following options are important for streaming models:
+
+  ``--decode-chunk-size``
+
+    For streaming models, we will calculate the chunk-wise attention, ``--decode-chunk-size``
+    indicates the chunk length (in frames after subsampling) for chunk-wise attention.
+    For ``real streaming decoding``, we will process ``decode-chunk-size`` acoustic frames at each time.
+
+  ``--left-context``
+
+    ``--left-context`` indicates how many left context frames (after subsampling) can be seen
+    for current chunk when calculating chunk-wise attention. Normally, ``left-context`` should equal
+    to ``decode-chunk-size * num-left-chunks``, where ``num-left-chunks`` is the option used
+    to train this model.
+
+  ``--num-decode-streams``
+
+    The number of decoding streams that can be run in parallel (very similar to the ``bath size``).
+    For ``real streaming decoding``, the batches will be packed dynamically, for example, if the
+    ``num-decode-streams`` equals to 10, then, sequence 1 to 10 will be decoded at first, after a while,
+    suppose sequence 1 and 2 are done, so, sequence 3 to 12 will be processed parallelly in a batch.
+
+
+.. NOTE::
+
+   We also try adding ``--right-context`` in the real streaming decoding, but it seems not to benefit
+   the performance for all the models, the reasons might be the training and decoding mismatch. You
+   can try decoding with ``--right-context`` to see if it helps. The default value is 0.
+
+
+The following shows two examples (for the two types of checkpoints):
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for epoch in 25 20; do
+      for avg in 7 5 3 1; do
+        ./pruned_transducer_stateless4/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --decode-chunk-size 16 \
+          --left-context 64 \
+          --num-decode-streams 100 \
+          --exp-dir pruned_transducer_stateless4/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for iter in 474000; do
+      for avg in 8 10 12 14 16 18; do
+        ./pruned_transducer_stateless4/decode.py \
+          --iter $iter \
+          --avg $avg \
+          --decode-chunk-size 16 \
+          --left-context 64 \
+          --num-decode-streams 100 \
+          --exp-dir pruned_transducer_stateless4/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. tip::
+
+  Supporting decoding methods are as follows:
+
+    - ``greedy_search`` : It takes the symbol with largest posterior probability
+      of each frame as the decoding result.
+
+    - ``beam_search`` :  It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf and
+      `espnet/nets/beam_search_transducer.py <https://github.com/espnet/espnet/blob/master/espnet/nets/beam_search_transducer.py#L247>`_
+      is used as a reference. Basicly, it keeps topk states for each frame, and expands the kept states with their own contexts to
+      next frame.
+
+    - ``modified_beam_search`` : It implements the same algorithm as ``beam_search`` above, but it
+      runs in batch mode with ``--max-sym-per-frame=1`` being hardcoded.
+
+    - ``fast_beam_search`` : It implements graph composition between the output ``log_probs`` and
+      given ``FSAs``. It is hard to describe the details in several lines of texts, you can read
+      our paper in https://arxiv.org/pdf/2211.00484.pdf or our `rnnt decode code in k2 <https://github.com/k2-fsa/k2/blob/master/k2/csrc/rnnt_decode.h>`_. ``fast_beam_search`` can decode with ``FSAs`` on GPU efficiently.
+
+    - ``fast_beam_search_LG`` : The same as ``fast_beam_search`` above, ``fast_beam_search`` uses
+      an trivial graph that has only one state, while ``fast_beam_search_LG`` uses an LG graph
+      (with N-gram LM).
+
+    - ``fast_beam_search_nbest`` : It produces the decoding results as follows:
+
+      - (1) Use ``fast_beam_search`` to get a lattice
+      - (2) Select ``num_paths`` paths from the lattice using ``k2.random_paths()``
+      - (3) Unique the selected paths
+      - (4) Intersect the selected paths with the lattice and compute the
+            shortest path from the intersection result
+      - (5) The path with the largest score is used as the decoding output.
+
+    - ``fast_beam_search_nbest_LG`` : It implements same logic as ``fast_beam_search_nbest``, the
+      only difference is that it uses ``fast_beam_search_LG`` to generate the lattice.
+
+.. NOTE::
+
+  The supporting decoding methods in ``streaming_decode.py`` might be less than that in ``decode.py``, if needed,
+  you can implement them by yourself or file a issue in `icefall <https://github.com/k2-fsa/icefall/issues>`_ .
+
+
+Export Model
+------------
+
+`pruned_transducer_stateless4/export.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless4/export.py>`_ supports exporting checkpoints from ``pruned_transducer_stateless4/exp`` in the following ways.
+
+Export ``model.state_dict()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Checkpoints saved by ``pruned_transducer_stateless4/train.py`` also include
+``optimizer.state_dict()``. It is useful for resuming training. But after training,
+we are interested only in ``model.state_dict()``. You can use the following
+command to extract ``model.state_dict()``.
+
+.. code-block:: bash
+
+  # Assume that --epoch 25 --avg 3 produces the smallest WER
+  # (You can get such information after running ./pruned_transducer_stateless4/decode.py)
+
+  epoch=25
+  avg=3
+
+  ./pruned_transducer_stateless4/export.py \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --streaming-model 1 \
+    --causal-convolution 1 \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch $epoch \
+    --avg  $avg
+
+.. caution::
+
+   ``--streaming-model`` and ``--causal-convolution`` require to be True to export
+   a streaming mdoel.
+
+It will generate a file ``./pruned_transducer_stateless4/exp/pretrained.pt``.
+
+.. hint::
+
+   To use the generated ``pretrained.pt`` for ``pruned_transducer_stateless4/decode.py``,
+   you can run:
+
+   .. code-block:: bash
+
+      cd pruned_transducer_stateless4/exp
+      ln -s pretrained.pt epoch-999.pt
+
+   And then pass ``--epoch 999 --avg 1 --use-averaged-model 0`` to
+   ``./pruned_transducer_stateless4/decode.py``.
+
+To use the exported model with ``./pruned_transducer_stateless4/pretrained.py``, you
+can run:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless4/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless4/exp/pretrained.pt \
+    --simulate-streaming 1 \
+    --causal-convolution 1 \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+
+Export model using ``torch.jit.script()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless4/export.py \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --streaming-model 1 \
+    --causal-convolution 1 \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 25 \
+    --avg 3 \
+    --jit 1
+
+.. caution::
+
+   ``--streaming-model`` and ``--causal-convolution`` require to be True to export
+   a streaming mdoel.
+
+It will generate a file ``cpu_jit.pt`` in the given ``exp_dir``. You can later
+load it by ``torch.jit.load("cpu_jit.pt")``.
+
+Note ``cpu`` in the name ``cpu_jit.pt`` means the parameters when loaded into Python
+are on CPU. You can use ``to("cuda")`` to move them to a CUDA device.
+
+.. NOTE::
+
+   You will need this ``cpu_jit.pt`` when deploying with Sherpa framework.
+
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following links:
+
+  - `pruned_transducer_stateless <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless_20220625>`_
+
+  - `pruned_transducer_stateless2 <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless2_20220625>`_
+
+  - `pruned_transducer_stateless4 <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless4_20220625>`_
+
+  - `pruned_transducer_stateless5 <https://huggingface.co/pkufool/icefall_librispeech_streaming_pruned_transducer_stateless5_20220729>`_
+
+  See `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  for the details of the above pretrained models
+
+
+Deploy with Sherpa
+------------------
+
+Please see `<https://k2-fsa.github.io/sherpa/python/streaming_asr/conformer/index.html#>`_
+for how to deploy the models in ``sherpa``.

docs/source/recipes/Streaming-ASR/librispeech/zipformer_transducer.rst

@@ -0,0 +1,654 @@
+Zipformer Transducer
+====================
+
+This tutorial shows you how to run a **streaming** zipformer transducer model
+with the `LibriSpeech <https://www.openslr.org/12>`_ dataset.
+
+.. Note::
+
+   The tutorial is suitable for `pruned_transducer_stateless7_streaming <https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless7_streaming>`_,
+
+.. HINT::
+
+  We assume you have read the page :ref:`install icefall` and have setup
+  the environment for ``icefall``.
+
+.. HINT::
+
+  We recommend you to use a GPU or several GPUs to run this recipe.
+
+.. hint::
+
+   Please scroll down to the bottom of this page to find download links
+   for pretrained models if you don't want to train a model from scratch.
+
+
+We use pruned RNN-T to compute the loss.
+
+.. note::
+
+   You can find the paper about pruned RNN-T at the following address:
+
+   `<https://arxiv.org/abs/2206.13236>`_
+
+The transducer model consists of 3 parts:
+
+  - Encoder, a.k.a, the transcription network. We use a Zipformer model (proposed by Daniel Povey)
+  - Decoder, a.k.a, the prediction network. We use a stateless model consisting of
+    ``nn.Embedding`` and ``nn.Conv1d``
+  - Joiner, a.k.a, the joint network.
+
+.. caution::
+
+   Contrary to the conventional RNN-T models, we use a stateless decoder.
+   That is, it has no recurrent connections.
+
+
+Data preparation
+----------------
+
+.. hint::
+
+   The data preparation is the same as other recipes on LibriSpeech dataset,
+   if you have finished this step, you can skip to ``Training`` directly.
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh
+
+The script ``./prepare.sh`` handles the data preparation for you, **automagically**.
+All you need to do is to run it.
+
+The data preparation contains several stages, you can use the following two
+options:
+
+  - ``--stage``
+  - ``--stop-stage``
+
+to control which stage(s) should be run. By default, all stages are executed.
+
+
+For example,
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./prepare.sh --stage 0 --stop-stage 0
+
+means to run only stage 0.
+
+To run stage 2 to stage 5, use:
+
+.. code-block:: bash
+
+  $ ./prepare.sh --stage 2 --stop-stage 5
+
+.. HINT::
+
+  If you have pre-downloaded the `LibriSpeech <https://www.openslr.org/12>`_
+  dataset and the `musan <http://www.openslr.org/17/>`_ dataset, say,
+  they are saved in ``/tmp/LibriSpeech`` and ``/tmp/musan``, you can modify
+  the ``dl_dir`` variable in ``./prepare.sh`` to point to ``/tmp`` so that
+  ``./prepare.sh`` won't re-download them.
+
+.. NOTE::
+
+  All generated files by ``./prepare.sh``, e.g., features, lexicon, etc,
+  are saved in ``./data`` directory.
+
+We provide the following YouTube video showing how to run ``./prepare.sh``.
+
+.. note::
+
+   To get the latest news of `next-gen Kaldi <https://github.com/k2-fsa>`_, please subscribe
+   the following YouTube channel by `Nadira Povey <https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_:
+
+      `<https://www.youtube.com/channel/UC_VaumpkmINz1pNkFXAN9mw>`_
+
+..  youtube:: ofEIoJL-mGM
+
+
+Training
+--------
+
+Configurable options
+~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless7_streaming/train.py --help
+
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--exp-dir``
+
+    The directory to save checkpoints, training logs and tensorboard.
+
+  - ``--full-libri``
+
+    If it's True, the training part uses all the training data, i.e.,
+    960 hours. Otherwise, the training part uses only the subset
+    ``train-clean-100``, which has 100 hours of training data.
+
+    .. CAUTION::
+      The training set is perturbed by speed with two factors: 0.9 and 1.1.
+      If ``--full-libri`` is True, each epoch actually processes
+      ``3x960 == 2880`` hours of data.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./pruned_transducer_stateless7_streaming/train.py --num-epochs 30`` trains for 30 epochs
+    and generates ``epoch-1.pt``, ``epoch-2.pt``, ..., ``epoch-30.pt``
+    in the folder ``./pruned_transducer_stateless7_streaming/exp``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./pruned_transducer_stateless7_streaming/train.py --start-epoch 10`` loads the
+    checkpoint ``./pruned_transducer_stateless7_streaming/exp/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./pruned_transducer_stateless7_streaming/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ ./pruned_transducer_stateless7_streaming/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/librispeech/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./pruned_transducer_stateless7_streaming/train.py --world-size 1
+
+    .. caution::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      Multi-GPU multi-machine DDP training will be added later.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+  - ``--use-fp16``
+
+    If it is True, the model will train with half precision, from our experiment
+    results, by using half precision you can train with two times larger ``--max-duration``
+    so as to get almost 2X speed up.
+
+    We recommend using ``--use-fp16 True``.
+
+  - ``--short-chunk-size``
+
+    When training a streaming attention model with chunk masking, the chunk size
+    would be either max sequence length of current batch or uniformly sampled from
+    (1, short_chunk_size). The default value is 50, you don't have to change it most of the time.
+
+  - ``--num-left-chunks``
+
+    It indicates how many left context (in chunks) that can be seen when calculating attention.
+    The default value is 4, you don't have to change it most of the time.
+
+
+  - ``--decode-chunk-len``
+
+    The chunk size for decoding (in frames before subsampling). It is used for validation.
+    The default value is 32 (i.e., 320ms).
+
+
+Pre-configured options
+~~~~~~~~~~~~~~~~~~~~~~
+
+There are some training options, e.g., number of encoder layers,
+encoder dimension, decoder dimension, number of warmup steps etc,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`pruned_transducer_stateless7_streaming/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py>`_
+
+You don't need to change these pre-configured parameters. If you really need to change
+them, please modify ``./pruned_transducer_stateless7_streaming/train.py`` directly.
+
+
+Training logs
+~~~~~~~~~~~~~
+
+Training logs and checkpoints are saved in ``--exp-dir`` (e.g. ``pruned_transducer_stateless7_streaming/exp``.
+You will find the following files in that directory:
+
+  - ``epoch-1.pt``, ``epoch-2.pt``, ...
+
+    These are checkpoint files saved at the end of each epoch, containing model
+    ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``epoch-10.pt``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless7_streaming/train.py --start-epoch 11
+
+  - ``checkpoint-436000.pt``, ``checkpoint-438000.pt``, ...
+
+    These are checkpoint files saved every ``--save-every-n`` batches,
+    containing model ``state_dict`` and optimizer ``state_dict``.
+    To resume training from some checkpoint, say ``checkpoint-436000``, you can use:
+
+      .. code-block:: bash
+
+        $ ./pruned_transducer_stateless7_streaming/train.py --start-batch 436000
+
+  - ``tensorboard/``
+
+    This folder contains tensorBoard logs. Training loss, validation loss, learning
+    rate, etc, are recorded in these logs. You can visualize them by:
+
+      .. code-block:: bash
+
+        $ cd pruned_transducer_stateless7_streaming/exp/tensorboard
+        $ tensorboard dev upload --logdir . --description "pruned transducer training for LibriSpeech with icefall"
+
+  .. hint::
+
+    If you don't have access to google, you can use the following command
+    to view the tensorboard log locally:
+
+      .. code-block:: bash
+
+        cd pruned_transducer_stateless7_streaming/exp/tensorboard
+        tensorboard --logdir . --port 6008
+
+    It will print the following message:
+
+      .. code-block::
+
+        Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
+        TensorBoard 2.8.0 at http://localhost:6008/ (Press CTRL+C to quit)
+
+    Now start your browser and go to `<http://localhost:6008>`_ to view the tensorboard
+    logs.
+
+
+  - ``log/log-train-xxxx``
+
+    It is the detailed training log in text format, same as the one
+    you saw printed to the console during training.
+
+Usage example
+~~~~~~~~~~~~~
+
+You can use the following command to start the training using 4 GPUs:
+
+.. code-block:: bash
+
+  export CUDA_VISIBLE_DEVICES="0,1,2,3"
+  ./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 \
+    --full-libri 1 \
+    --max-duration 550
+
+Decoding
+--------
+
+The decoding part uses checkpoints saved by the training part, so you have
+to run the training part first.
+
+.. hint::
+
+   There are two kinds of checkpoints:
+
+    - (1) ``epoch-1.pt``, ``epoch-2.pt``, ..., which are saved at the end
+      of each epoch. You can pass ``--epoch`` to
+      ``pruned_transducer_stateless7_streaming/decode.py`` to use them.
+
+    - (2) ``checkpoints-436000.pt``, ``epoch-438000.pt``, ..., which are saved
+      every ``--save-every-n`` batches. You can pass ``--iter`` to
+      ``pruned_transducer_stateless7_streaming/decode.py`` to use them.
+
+    We suggest that you try both types of checkpoints and choose the one
+    that produces the lowest WERs.
+
+.. tip::
+
+    To decode a streaming model, you can use either ``simulate streaming decoding`` in ``decode.py`` or
+    ``real chunk-wise streaming decoding`` in ``streaming_decode.py``. The difference between ``decode.py`` and
+    ``streaming_decode.py`` is that, ``decode.py`` processes the whole acoustic frames at one time with masking (i.e. same as training),
+    but ``streaming_decode.py`` processes the acoustic frames chunk by chunk.
+
+.. NOTE::
+
+   ``simulate streaming decoding`` in ``decode.py`` and ``real chunk-size streaming decoding`` in ``streaming_decode.py`` should
+   produce almost the same results given the same ``--decode-chunk-len``.
+
+
+Simulate streaming decoding
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless7_streaming/decode.py --help
+
+shows the options for decoding.
+The following options are important for streaming models:
+
+  ``--decode-chunk-len``
+
+    It is same as in ``train.py``, which specifies the chunk size for decoding (in frames before subsampling).
+    The default value is 32 (i.e., 320ms).
+
+
+The following shows two examples (for the two types of checkpoints):
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for epoch in 30; do
+      for avg in 12 11 10 9 8; do
+        ./pruned_transducer_stateless7_streaming/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --decode-chunk-len 32 \
+          --exp-dir pruned_transducer_stateless7_streaming/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for iter in 474000; do
+      for avg in 8 10 12 14 16 18; do
+        ./pruned_transducer_stateless7_streaming/decode.py \
+          --iter $iter \
+          --avg $avg \
+          --decode-chunk-len 32 \
+          --exp-dir pruned_transducer_stateless7_streaming/exp \
+          --max-duration 600 \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+Real streaming decoding
+~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  $ cd egs/librispeech/ASR
+  $ ./pruned_transducer_stateless7_streaming/streaming_decode.py --help
+
+shows the options for decoding.
+The following options are important for streaming models:
+
+  ``--decode-chunk-len``
+
+    It is same as in ``train.py``, which specifies the chunk size for decoding (in frames before subsampling).
+    The default value is 32 (i.e., 320ms).
+    For ``real streaming decoding``, we will process ``decode-chunk-len`` acoustic frames at each time.
+
+  ``--num-decode-streams``
+
+    The number of decoding streams that can be run in parallel (very similar to the ``bath size``).
+    For ``real streaming decoding``, the batches will be packed dynamically, for example, if the
+    ``num-decode-streams`` equals to 10, then, sequence 1 to 10 will be decoded at first, after a while,
+    suppose sequence 1 and 2 are done, so, sequence 3 to 12 will be processed parallelly in a batch.
+
+
+The following shows two examples (for the two types of checkpoints):
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for epoch in 30; do
+      for avg in 12 11 10 9 8; do
+        ./pruned_transducer_stateless7_streaming/decode.py \
+          --epoch $epoch \
+          --avg $avg \
+          --decode-chunk-len 32 \
+          --num-decode-streams 100 \
+          --exp-dir pruned_transducer_stateless7_streaming/exp \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. code-block:: bash
+
+  for m in greedy_search fast_beam_search modified_beam_search; do
+    for iter in 474000; do
+      for avg in 8 10 12 14 16 18; do
+        ./pruned_transducer_stateless7_streaming/decode.py \
+          --iter $iter \
+          --avg $avg \
+          --decode-chunk-len 16 \
+          --num-decode-streams 100 \
+          --exp-dir pruned_transducer_stateless7_streaming/exp \
+          --decoding-method $m
+      done
+    done
+  done
+
+
+.. tip::
+
+  Supporting decoding methods are as follows:
+
+    - ``greedy_search`` : It takes the symbol with largest posterior probability
+      of each frame as the decoding result.
+
+    - ``beam_search`` :  It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf and
+      `espnet/nets/beam_search_transducer.py <https://github.com/espnet/espnet/blob/master/espnet/nets/beam_search_transducer.py#L247>`_
+      is used as a reference. Basicly, it keeps topk states for each frame, and expands the kept states with their own contexts to
+      next frame.
+
+    - ``modified_beam_search`` : It implements the same algorithm as ``beam_search`` above, but it
+      runs in batch mode with ``--max-sym-per-frame=1`` being hardcoded.
+
+    - ``fast_beam_search`` : It implements graph composition between the output ``log_probs`` and
+      given ``FSAs``. It is hard to describe the details in several lines of texts, you can read
+      our paper in https://arxiv.org/pdf/2211.00484.pdf or our `rnnt decode code in k2 <https://github.com/k2-fsa/k2/blob/master/k2/csrc/rnnt_decode.h>`_. ``fast_beam_search`` can decode with ``FSAs`` on GPU efficiently.
+
+    - ``fast_beam_search_LG`` : The same as ``fast_beam_search`` above, ``fast_beam_search`` uses
+      an trivial graph that has only one state, while ``fast_beam_search_LG`` uses an LG graph
+      (with N-gram LM).
+
+    - ``fast_beam_search_nbest`` : It produces the decoding results as follows:
+
+      - (1) Use ``fast_beam_search`` to get a lattice
+      - (2) Select ``num_paths`` paths from the lattice using ``k2.random_paths()``
+      - (3) Unique the selected paths
+      - (4) Intersect the selected paths with the lattice and compute the
+            shortest path from the intersection result
+      - (5) The path with the largest score is used as the decoding output.
+
+    - ``fast_beam_search_nbest_LG`` : It implements same logic as ``fast_beam_search_nbest``, the
+      only difference is that it uses ``fast_beam_search_LG`` to generate the lattice.
+
+.. NOTE::
+
+  The supporting decoding methods in ``streaming_decode.py`` might be less than that in ``decode.py``, if needed,
+  you can implement them by yourself or file a issue in `icefall <https://github.com/k2-fsa/icefall/issues>`_ .
+
+
+Export Model
+------------
+
+Currently it supports exporting checkpoints from ``pruned_transducer_stateless7_streaming/exp`` in the following ways.
+
+Export ``model.state_dict()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Checkpoints saved by ``pruned_transducer_stateless7_streaming/train.py`` also include
+``optimizer.state_dict()``. It is useful for resuming training. But after training,
+we are interested only in ``model.state_dict()``. You can use the following
+command to extract ``model.state_dict()``.
+
+.. code-block:: bash
+
+  # Assume that --epoch 30 --avg 9 produces the smallest WER
+  # (You can get such information after running ./pruned_transducer_stateless7_streaming/decode.py)
+
+  epoch=30
+  avg=9
+
+  ./pruned_transducer_stateless7_streaming/export.py \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch $epoch \
+    --avg  $avg \
+    --use-averaged-model=True \
+    --decode-chunk-len 32
+
+It will generate a file ``./pruned_transducer_stateless7_streaming/exp/pretrained.pt``.
+
+.. hint::
+
+   To use the generated ``pretrained.pt`` for ``pruned_transducer_stateless7_streaming/decode.py``,
+   you can run:
+
+   .. code-block:: bash
+
+      cd pruned_transducer_stateless7_streaming/exp
+      ln -s pretrained.pt epoch-999.pt
+
+   And then pass ``--epoch 999 --avg 1 --use-averaged-model 0`` to
+   ``./pruned_transducer_stateless7_streaming/decode.py``.
+
+To use the exported model with ``./pruned_transducer_stateless7_streaming/pretrained.py``, you
+can run:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_streaming/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless7_streaming/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    --decode-chunk-len 32 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+
+Export model using ``torch.jit.script()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_streaming/export.py \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --epoch 30 \
+    --avg 9 \
+    --decode-chunk-len 32 \
+    --jit 1
+
+.. caution::
+
+   ``--decode-chunk-len`` is required to export a ScriptModule.
+
+It will generate a file ``cpu_jit.pt`` in the given ``exp_dir``. You can later
+load it by ``torch.jit.load("cpu_jit.pt")``.
+
+Note ``cpu`` in the name ``cpu_jit.pt`` means the parameters when loaded into Python
+are on CPU. You can use ``to("cuda")`` to move them to a CUDA device.
+
+Export model using ``torch.jit.trace()``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: bash
+
+  epoch=30
+  avg=9
+
+  ./pruned_transducer_stateless7_streaming/jit_trace_export.py \
+    --bpe-model data/lang_bpe_500/bpe.model \
+    --use-averaged-model=True \
+    --decode-chunk-len 32 \
+    --exp-dir ./pruned_transducer_stateless7_streaming/exp \
+    --epoch $epoch \
+    --avg $avg
+
+.. caution::
+
+   ``--decode-chunk-len`` is required to export a ScriptModule.
+
+It will generate 3 files:
+
+  - ``./pruned_transducer_stateless7_streaming/exp/encoder_jit_trace.pt``
+  - ``./pruned_transducer_stateless7_streaming/exp/decoder_jit_trace.pt``
+  - ``./pruned_transducer_stateless7_streaming/exp/joiner_jit_trace.pt``
+
+To use the generated files with ``./pruned_transducer_stateless7_streaming/jit_trace_pretrained.py``:
+
+.. code-block:: bash
+
+  ./pruned_transducer_stateless7_streaming/jit_trace_pretrained.py \
+    --encoder-model-filename ./pruned_transducer_stateless7_streaming/exp/encoder_jit_trace.pt \
+    --decoder-model-filename ./pruned_transducer_stateless7_streaming/exp/decoder_jit_trace.pt \
+    --joiner-model-filename ./pruned_transducer_stateless7_streaming/exp/joiner_jit_trace.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --decode-chunk-len 32 \
+    /path/to/foo.wav
+
+
+Download pretrained models
+--------------------------
+
+If you don't want to train from scratch, you can download the pretrained models
+by visiting the following links:
+
+  - `pruned_transducer_stateless7_streaming <https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29>`_
+
+  See `<https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md>`_
+  for the details of the above pretrained models
+
+Deploy with Sherpa
+------------------
+
+Please see `<https://k2-fsa.github.io/sherpa/python/streaming_asr/conformer/index.html#>`_
+for how to deploy the models in ``sherpa``.

docs/source/recipes/index.rst

@@ -13,7 +13,5 @@ We may add recipes for other tasks as well in the future.
    :maxdepth: 2
    :caption: Table of Contents
 
-   aishell/index
-   librispeech/index
-   timit/index
-   yesno/index
+   Non-streaming-ASR/index
+   Streaming-ASR/index

egs/aidatatang_200zh/ASR/local/compute_fbank_aidatatang_200zh.py

@@ -87,9 +87,7 @@ def compute_fbank_aidatatang_200zh(num_mel_bins: int = 80):
             )
             if "train" in partition:
                 cut_set = (
-                    cut_set
-                    + cut_set.perturb_speed(0.9)
-                    + cut_set.perturb_speed(1.1)
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
                 )
             cut_set = cut_set.compute_and_store_features(
                 extractor=extractor,
@@ -116,9 +114,7 @@ def get_args():
 
 
 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)
 

egs/aidatatang_200zh/ASR/local/prepare_char.py

@@ -86,9 +86,7 @@ def lexicon_to_fst_no_sil(
         cur_state = loop_state
 
         word = word2id[word]
-        pieces = [
-            token2id[i] if i in token2id else token2id["<unk>"] for i in pieces
-        ]
+        pieces = [token2id[i] if i in token2id else token2id["<unk>"] for i in pieces]
 
         for i in range(len(pieces) - 1):
             w = word if i == 0 else eps
@@ -142,9 +140,7 @@ def contain_oov(token_sym_table: Dict[str, int], tokens: List[str]) -> bool:
     return False
 
 
-def generate_lexicon(
-    token_sym_table: Dict[str, int], words: List[str]
-) -> Lexicon:
+def generate_lexicon(token_sym_table: Dict[str, int], words: List[str]) -> Lexicon:
     """Generate a lexicon from a word list and token_sym_table.
 
     Args:

egs/aidatatang_200zh/ASR/local/prepare_lang.py

@@ -317,9 +317,7 @@ def lexicon_to_fst(
 
 def get_args():
     parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--lang-dir", type=str, help="The lang dir, data/lang_phone"
-    )
+    parser.add_argument("--lang-dir", type=str, help="The lang dir, data/lang_phone")
     return parser.parse_args()
 
 

egs/aidatatang_200zh/ASR/local/test_prepare_lang.py

@@ -88,9 +88,7 @@ def test_read_lexicon(filename: str):
     fsa.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
     fsa.draw("L.pdf", title="L")
 
-    fsa_disambig = lexicon_to_fst(
-        lexicon_disambig, phone2id=phone2id, word2id=word2id
-    )
+    fsa_disambig = lexicon_to_fst(lexicon_disambig, phone2id=phone2id, word2id=word2id)
     fsa_disambig.labels_sym = k2.SymbolTable.from_file("phones.txt")
     fsa_disambig.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
     fsa_disambig.draw("L_disambig.pdf", title="L_disambig")

egs/aidatatang_200zh/ASR/local/text2token.py

@@ -56,9 +56,7 @@ def get_parser():
     parser.add_argument(
         "--skip-ncols", "-s", default=0, type=int, help="skip first n columns"
     )
-    parser.add_argument(
-        "--space", default="<space>", type=str, help="space symbol"
-    )
+    parser.add_argument("--space", default="<space>", type=str, help="space symbol")
     parser.add_argument(
         "--non-lang-syms",
         "-l",
@@ -66,9 +64,7 @@ def get_parser():
         type=str,
         help="list of non-linguistic symobles, e.g., <NOISE> etc.",
     )
-    parser.add_argument(
-        "text", type=str, default=False, nargs="?", help="input text"
-    )
+    parser.add_argument("text", type=str, default=False, nargs="?", help="input text")
     parser.add_argument(
         "--trans_type",
         "-t",
@@ -108,8 +104,7 @@ def token2id(
             if token_type == "lazy_pinyin":
                 text = lazy_pinyin(chars_list)
                 sub_ids = [
-                    token_table[txt] if txt in token_table else oov_id
-                    for txt in text
+                    token_table[txt] if txt in token_table else oov_id for txt in text
                 ]
                 ids.append(sub_ids)
             else:  # token_type = "pinyin"
@@ -135,9 +130,7 @@ def main():
     if args.text:
         f = codecs.open(args.text, encoding="utf-8")
     else:
-        f = codecs.getreader("utf-8")(
-            sys.stdin if is_python2 else sys.stdin.buffer
-        )
+        f = codecs.getreader("utf-8")(sys.stdin if is_python2 else sys.stdin.buffer)
 
     sys.stdout = codecs.getwriter("utf-8")(
         sys.stdout if is_python2 else sys.stdout.buffer

egs/aidatatang_200zh/ASR/prepare.sh

@@ -1,5 +1,8 @@
 #!/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
 
 stage=-1
@@ -113,4 +116,3 @@ if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
     ./local/prepare_char.py
   fi
 fi
-