a bilingual recipe similar to the multi-zh_hans (#1265)

.github/scripts/run-multi-zh_hans-zipformer.sh → .github/scripts/run-multi-corpora-zipformer.sh

@@ -95,3 +95,41 @@ for method in modified_beam_search fast_beam_search; do
   $repo/test_wavs/DEV_T0000000001.wav \
   $repo/test_wavs/DEV_T0000000002.wav
 done
+
+rm -rf $repo
+
+cd ../../../egs/multi_zh_en/ASR
+log "==== Test icefall-asr-zipformer-multi-zh-en-2023-11-22 ===="
+repo_url=https://huggingface.co/zrjin/icefall-asr-zipformer-multi-zh-en-2023-11-22/
+
+log "Downloading pre-trained model from $repo_url"
+git lfs install
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+./zipformer/pretrained.py \
+  --checkpoint $repo/exp/pretrained.pt \
+  --bpe-model $repo/data/lang_bbpe_2000/bbpe.model \
+  --method greedy_search \
+$repo/test_wavs/_1634_210_2577_1_1525157964032_3712259_29.wav \
+$repo/test_wavs/_1634_210_2577_1_1525157964032_3712259_55.wav \
+$repo/test_wavs/_1634_210_2577_1_1525157964032_3712259_75.wav
+
+for method in modified_beam_search fast_beam_search; do
+  log "$method"
+
+  ./zipformer/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --bpe-model $repo/data/lang_bbpe_2000/bbpe.model \
+  $repo/test_wavs/_1634_210_2577_1_1525157964032_3712259_29.wav \
+  $repo/test_wavs/_1634_210_2577_1_1525157964032_3712259_55.wav \
+  $repo/test_wavs/_1634_210_2577_1_1525157964032_3712259_75.wav
+done
+
+rm -rf $repo

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

@@ -14,7 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-name: run-multi-zh_hans-zipformer
+name: run-multi-corpora-zipformer
 
 on:
   push:
@@ -24,12 +24,12 @@ on:
     types: [labeled]
 
 concurrency:
-  group: run_multi-zh_hans_zipformer-${{ github.ref }}
+  group: run_multi-corpora_zipformer-${{ github.ref }}
   cancel-in-progress: true
 
 jobs:
-  run_multi-zh_hans_zipformer:
-    if: github.event.label.name == 'onnx' || github.event.label.name == 'ready' || github.event_name == 'push' || github.event.label.name == 'multi-zh_hans' || github.event.label.name == 'zipformer'
+  run_multi-corpora_zipformer:
+    if: github.event.label.name == 'onnx' || github.event.label.name == 'ready' || github.event_name == 'push' || github.event.label.name == 'multi-zh_hans' || github.event.label.name == 'zipformer' || github.event.label.name == 'multi-corpora'
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
@@ -81,4 +81,4 @@ jobs:
           export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
           export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
 
-          .github/scripts/run-multi-zh_hans-zipformer.sh
+          .github/scripts/run-multi-corpora-zipformer.sh

egs/multi_zh_en/ASR/README.md

@@ -0,0 +1,19 @@
+# Introduction
+
+This recipe includes scripts for training Zipformer model using both English and Chinese datasets.
+
+# Included Training Sets
+
+1. LibriSpeech (English)
+2. AiShell-2 (Chinese)
+3. TAL-CSASR (Code-Switching, Chinese and English)
+
+|Datset| Number of hours| URL|
+|---|---:|---|
+|**TOTAL**|2,547|---|
+|LibriSpeech|960|https://www.openslr.org/12/|
+|AiShell-2|1,000|http://www.aishelltech.com/aishell_2|
+|TAL-CSASR|587|https://ai.100tal.com/openData/voice|
+
+
+

egs/multi_zh_en/ASR/RESULTS.md

@@ -0,0 +1,44 @@
+## Results
+
+### Zh-En datasets bpe-based training results (Non-streaming) on Zipformer model
+
+This is the [pull request #1238](https://github.com/k2-fsa/icefall/pull/1265) in icefall.
+
+#### Non-streaming (Byte-Level BPE vocab_size=2000)
+
+Best results (num of params : ~69M):
+
+The training command:
+
+```
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 35 \
+  --use-fp16 1 \
+  --max-duration 1000 \
+  --num-workers 8
+```
+
+The decoding command:
+
+```
+for method in greedy_search modified_beam_search fast_beam_search; do
+    ./zipformer/decode.py \
+    --epoch 34 \
+    --avg 19 \
+    --decoding-method $method
+done
+```
+
+Word Error Rates (WERs) listed below are produced by the checkpoint of the 20th epoch using greedy search and BPE model (# tokens is 2000).
+
+|       Datasets       | TAL-CSASR | TAL-CSASR | AiShell-2 | AiShell-2 | LibriSpeech | LibriSpeech | 
+|----------------------|-----------|-----------|-----------|-----------|-------------|-------------|
+|   Zipformer WER (%)  |    dev    |   test    |    dev    |   test    |  test-clean |  test-other | 
+|     greedy_search    |   6.65    |   6.69    |    6.57   |   7.03    |    2.43     |    5.70     |
+| modified_beam_search |   6.46    |   6.51    |    6.18   |   6.60    |    2.41     |    5.57     |
+|   fast_beam_search   |   6.57    |   6.68    |    6.40   |   6.74    |    2.40     |    5.56     |
+
+Pre-trained model can be found here : https://huggingface.co/zrjin/icefall-asr-zipformer-multi-zh-en-2023-11-22, which is trained on LibriSpeech 960-hour training set (with speed perturbation), TAL-CSASR training set (with speed perturbation) and AiShell-2 (w/o speed perturbation).
+
+

egs/multi_zh_en/ASR/local/compile_lg.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compile_lg.py

egs/multi_zh_en/ASR/local/prepare_char.py

@@ -0,0 +1 @@
+../../../aishell/ASR/local/prepare_char.py

egs/multi_zh_en/ASR/local/prepare_for_bpe_model.py

@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Zengrui Jin)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script tokenizes the training transcript by CJK characters
+# and saves the result to transcript_chars.txt, which is used
+# to train the BPE model later.
+
+import argparse
+from pathlib import Path
+
+from tqdm.auto import tqdm
+
+from icefall.utils import tokenize_by_CJK_char
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Output directory.
+        The generated transcript_chars.txt is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--text",
+        type=str,
+        help="Training transcript.",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    text = Path(args.text)
+
+    assert lang_dir.exists() and text.exists(), f"{lang_dir} or {text} does not exist!"
+
+    transcript_path = lang_dir / "transcript_chars.txt"
+
+    with open(text, "r", encoding="utf-8") as fin:
+        with open(transcript_path, "w+", encoding="utf-8") as fout:
+            for line in tqdm(fin):
+                fout.write(tokenize_by_CJK_char(line) + "\n")
+
+
+if __name__ == "__main__":
+    main()

egs/multi_zh_en/ASR/local/prepare_lang.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang.py

egs/multi_zh_en/ASR/local/prepare_lang_bbpe.py

@@ -0,0 +1 @@
+../../../aishell/ASR/local/prepare_lang_bbpe.py

egs/multi_zh_en/ASR/local/prepare_lang_bpe.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/prepare_lang_bpe.py

egs/multi_zh_en/ASR/local/prepare_words.py

@@ -0,0 +1 @@
+../../../aishell2/ASR/local/prepare_words.py

egs/multi_zh_en/ASR/local/text2segments.py

@@ -0,0 +1 @@
+../../../wenetspeech/ASR/local/text2segments.py

egs/multi_zh_en/ASR/local/text2token.py

@@ -0,0 +1 @@
+../../../wenetspeech/ASR/local/text2token.py

egs/multi_zh_en/ASR/local/train_bbpe_model.py

@@ -0,0 +1 @@
+../../../aishell/ASR/local/train_bbpe_model.py

egs/multi_zh_en/ASR/local/validate_bpe_lexicon.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/validate_bpe_lexicon.py

egs/multi_zh_en/ASR/prepare.sh

@@ -0,0 +1,149 @@
+#!/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
+stop_stage=100
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+vocab_sizes=(
+  2000
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+log "Dataset: musan"
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Soft link fbank of musan"
+  mkdir -p data/fbank
+  if [ -e ../../librispeech/ASR/data/fbank/.musan.done ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/musan_feats) .
+    ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/musan_cuts.jsonl.gz) .
+    cd ../..
+  else
+    log "Abort! Please run ../../librispeech/ASR/prepare.sh --stage 4 --stop-stage 4"
+    exit 1
+  fi
+fi
+
+log "Dataset: LibriSpeech"
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Soft link fbank of LibriSpeech"
+  mkdir -p data/fbank
+  if [ -e ../../librispeech/ASR/data/fbank/.librispeech.done ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/librispeech_cuts*) .
+    ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/librispeech_feats*) .
+    cd ../..
+  else
+    log "Abort! Please run ../../librispeech/ASR/prepare.sh --stage 3 --stop-stage 3"
+    exit 1
+  fi
+fi
+
+log "Dataset: AiShell-2"
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Soft link fbank of AiShell-2"
+  mkdir -p data/fbank
+  if [ -e ../../aishell2/ASR/data/fbank/.aishell2.done ]; then
+    cd data/fbank
+    ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_cuts*) .
+    ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_feats*) .
+    cd ../..
+  else
+    log "Abort! Please run ../../aishell2/ASR/prepare.sh --stage 3 --stop-stage 3"
+    exit 1
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Prepare Byte BPE based lang"
+  mkdir -p data/fbank
+  if [ ! -d ../../aishell2/ASR/data/lang_char ] && [ ! -d ./data/lang_char ]; then
+    log "Abort! Please run ../../aishell2/ASR/prepare.sh --stage 3 --stop-stage 3"
+    exit 1
+  fi
+
+  if [ ! -d ../../librispeech/ASR/data/lang_bpe_500 ] && [ ! -d ./data/lang_bpe_500 ]; then
+    log "Abort! Please run ../../librispeech/ASR/prepare.sh --stage 6 --stop-stage 6"
+    exit 1
+  fi
+
+  cd data/
+  if [ ! -d ./lang_char ]; then
+    ln -svf $(realpath ../../../aishell2/ASR/data/lang_char) .
+  fi
+  if [ ! -d ./lang_bpe_500 ]; then
+    ln -svf $(realpath ../../../librispeech/ASR/data/lang_bpe_500) .
+  fi
+  cd ../
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bbpe_${vocab_size}
+    mkdir -p $lang_dir
+
+    cat data/lang_char/text data/lang_bpe_500/transcript_words.txt \
+      > $lang_dir/text
+
+    if [ ! -f $lang_dir/transcript_chars.txt ]; then
+      ./local/prepare_for_bpe_model.py \
+        --lang-dir ./$lang_dir \
+        --text $lang_dir/text
+    fi
+
+    if [ ! -f $lang_dir/text_words_segmentation ]; then
+      python3 ./local/text2segments.py \
+        --input-file ./data/lang_char/text \
+        --output-file $lang_dir/text_words_segmentation
+
+      cat ./data/lang_bpe_500/transcript_words.txt \
+        >> $lang_dir/text_words_segmentation
+
+      cat ./data/lang_char/text \
+        >> $lang_dir/text
+    fi
+
+    cat $lang_dir/text_words_segmentation | sed 's/ /\n/g' \
+      | sort -u | sed '/^$/d' | uniq > $lang_dir/words_no_ids.txt
+
+    if [ ! -f $lang_dir/words.txt ]; then
+      python3 ./local/prepare_words.py \
+        --input-file $lang_dir/words_no_ids.txt \
+        --output-file $lang_dir/words.txt
+    fi
+
+    if [ ! -f $lang_dir/bbpe.model ]; then
+      ./local/train_bbpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --transcript $lang_dir/text
+    fi
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bbpe.py --lang-dir $lang_dir
+
+      log "Validating $lang_dir/lexicon.txt"
+      ./local/validate_bpe_lexicon.py \
+        --lexicon $lang_dir/lexicon.txt \
+        --bpe-model $lang_dir/bbpe.model
+    fi
+  done 
+fi
+

egs/multi_zh_en/ASR/shared

@@ -0,0 +1 @@
+../../../icefall/shared