Add HLG decoding with OpenFst on CPU for aishell conformer_ctc (#1279)

.github/scripts/run-pre-trained-conformer-ctc.sh

@@ -8,7 +8,7 @@ log() {
   echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
 }
 
-cd egs/librispeech/ASR
+pushd egs/librispeech/ASR
 
 # repo_url=https://github.com/csukuangfj/icefall-asr-conformer-ctc-bpe-500
 repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09
@@ -112,3 +112,81 @@ log "Decoding with HLG on CPU with OpenFst"
   $repo/test_wavs/1089-134686-0001.wav \
   $repo/test_wavs/1221-135766-0001.wav \
   $repo/test_wavs/1221-135766-0002.wav
+
+rm -rf $repo
+
+popd
+
+log "Test aishell"
+
+pushd egs/aishell/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall_asr_aishell_conformer_ctc
+log "Downloading pre-trained model from $repo_url"
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
+repo=$(basename $repo_url)
+pushd $repo
+
+git lfs pull --include "exp/pretrained.pt"
+git lfs pull --include "data/lm/G_3_gram_char.fst.txt"
+
+popd
+
+log "Display test files"
+tree $repo/
+ls -lh $repo/test_wavs/*.wav
+
+log "CTC decoding"
+
+log "Exporting model with torchscript"
+
+pushd $repo/exp
+ln -s pretrained.pt epoch-99.pt
+popd
+
+./conformer_ctc/export.py \
+  --epoch 99 \
+  --avg 1 \
+  --exp-dir $repo/exp \
+  --tokens $repo/data/lang_char/tokens.txt \
+  --jit 1
+
+ls -lh $repo/exp
+
+log "Generating H.fst, HL.fst"
+
+./local/prepare_lang_fst.py  --lang-dir $repo/data/lang_char --ngram-G $repo/data/lm/G_3_gram_char.fst.txt
+
+ls -lh $repo/data/lang_char
+
+log "Decoding with H on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_H.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --H $repo/data/lang_char/H.fst \
+  --tokens $repo/data/lang_char/tokens.txt \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+log "Decoding with HL on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_HL.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --HL $repo/data/lang_char/HL.fst \
+  --words $repo/data/lang_char/words.txt \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+log "Decoding with HLG on CPU with OpenFst"
+
+./conformer_ctc/jit_pretrained_decode_with_HLG.py \
+  --nn-model $repo/exp/cpu_jit.pt \
+  --HLG $repo/data/lang_char/HLG.fst \
+  --words $repo/data/lang_char/words.txt \
+  $repo/test_wavs/0.wav \
+  $repo/test_wavs/1.wav \
+  $repo/test_wavs/2.wav
+
+rm -rf $repo

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

@@ -60,7 +60,7 @@ jobs:
 
       - name: Install Python dependencies
         run: |
-          grep -v '^#' ./requirements-ci.txt  | grep -v kaldifst | xargs -n 1 -L 1 pip install
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
           pip uninstall -y protobuf
           pip install --no-binary protobuf protobuf==3.20.*
 

egs/aishell/ASR/conformer_ctc/export.py

@@ -23,12 +23,12 @@
 import logging
 from pathlib import Path
 
+import k2
 import torch
 from conformer import Conformer
 
 from icefall.checkpoint import average_checkpoints, load_checkpoint
-from icefall.lexicon import Lexicon
-from icefall.utils import AttributeDict, str2bool
+from icefall.utils import AttributeDict, num_tokens, str2bool
 
 
 def get_parser():
@@ -63,11 +63,10 @@ def get_parser():
     )
 
     parser.add_argument(
-        "--lang-dir",
+        "--tokens",
         type=str,
-        default="data/lang_char",
-        help="""It contains language related input files such as "lexicon.txt"
-        """,
+        required=True,
+        help="Path to the tokens.txt.",
     )
 
     parser.add_argument(
@@ -98,16 +97,16 @@ def get_params() -> AttributeDict:
 def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
-    args.lang_dir = Path(args.lang_dir)
 
     params = get_params()
     params.update(vars(args))
 
-    logging.info(params)
+    # Load tokens.txt here
+    token_table = k2.SymbolTable.from_file(params.tokens)
 
-    lexicon = Lexicon(params.lang_dir)
-    max_token_id = max(lexicon.tokens)
-    num_classes = max_token_id + 1  # +1 for the blank
+    num_classes = num_tokens(token_table) + 1  # +1 for the blank
+
+    logging.info(params)
 
     device = torch.device("cpu")
     if torch.cuda.is_available():

egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_H.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_H.py

egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py

egs/aishell/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py

egs/aishell/ASR/local/prepare_lang_fst.py

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

egs/aishell/ASR/prepare.sh

@@ -143,6 +143,7 @@ if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
     ./local/prepare_lang.py --lang-dir $lang_phone_dir
   fi
 
+
   # Train a bigram P for MMI training
   if [ ! -f $lang_phone_dir/transcript_words.txt ]; then
     log "Generate data to train phone based bigram P"
@@ -203,6 +204,10 @@ if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
   if [ ! -f $lang_char_dir/L_disambig.pt ]; then
     ./local/prepare_char.py --lang-dir $lang_char_dir
   fi
+
+  if [ ! -f $lang_char_dir/HLG.fst ]; then
+    ./local/prepare_lang_fst.py  --lang-dir $lang_phone_dir --ngram-G ./data/lm/G_3_gram.fst.txt
+  fi
 fi
 
 if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then

egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_H.py

@@ -7,13 +7,26 @@
 
 Usage:
 
+(1) LibriSpeech conformer_ctc
+
     ./conformer_ctc/jit_pretrained_decode_with_H.py \
       --nn-model ./conformer_ctc/exp/cpu_jit.pt \
       --H ./data/lang_bpe_500/H.fst \
       --tokens ./data/lang_bpe_500/tokens.txt \
       ./download/LibriSpeech/test-clean/1089/134686/1089-134686-0002.flac \
       ./download/LibriSpeech/test-clean/1221/135766/1221-135766-0001.flac
 
+
+(2) AIShell conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_H.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --H ./data/lang_char/H.fst \
+      --tokens ./data/lang_char/tokens.txt \
+      ./BAC009S0764W0121.wav \
+      ./BAC009S0764W0122.wav \
+      ./BAC009S0764W0123.wav
+
 Note that to generate ./conformer_ctc/exp/cpu_jit.pt,
 you can use ./export.py --jit 1
 """
@@ -23,12 +36,11 @@
 import math
 from typing import Dict, List
 
-import kaldi_hmm_gmm
 import kaldifeat
 import kaldifst
 import torch
 import torchaudio
-from kaldi_hmm_gmm import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
 from torch.nn.utils.rnn import pad_sequence
 
 

egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HL.py

@@ -7,13 +7,26 @@
 
 Usage:
 
+(1) LibriSpeech conformer_ctc
+
     ./conformer_ctc/jit_pretrained_decode_with_HL.py \
       --nn-model ./conformer_ctc/exp/cpu_jit.pt \
       --HL ./data/lang_bpe_500/HL.fst \
       --words ./data/lang_bpe_500/words.txt \
       ./download/LibriSpeech/test-clean/1089/134686/1089-134686-0002.flac \
       ./download/LibriSpeech/test-clean/1221/135766/1221-135766-0001.flac
 
+(2) AIShell conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_HL.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --HL ./data/lang_char/HL.fst \
+      --words ./data/lang_char/words.txt \
+      ./BAC009S0764W0121.wav \
+      ./BAC009S0764W0122.wav \
+      ./BAC009S0764W0123.wav
+
+
 Note that to generate ./conformer_ctc/exp/cpu_jit.pt,
 you can use ./export.py --jit 1
 """
@@ -23,12 +36,11 @@
 import math
 from typing import Dict, List
 
-import kaldi_hmm_gmm
 import kaldifeat
 import kaldifst
 import torch
 import torchaudio
-from kaldi_hmm_gmm import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
 from torch.nn.utils.rnn import pad_sequence
 
 

egs/librispeech/ASR/conformer_ctc/jit_pretrained_decode_with_HLG.py

@@ -7,13 +7,25 @@
 
 Usage:
 
+(1) LibriSpeech conformer_ctc
+
     ./conformer_ctc/jit_pretrained_decode_with_HLG.py \
       --nn-model ./conformer_ctc/exp/cpu_jit.pt \
       --HLG ./data/lang_bpe_500/HLG.fst \
       --words ./data/lang_bpe_500/words.txt \
       ./download/LibriSpeech/test-clean/1089/134686/1089-134686-0002.flac \
       ./download/LibriSpeech/test-clean/1221/135766/1221-135766-0001.flac
 
+(2) AIShell conformer_ctc
+
+    ./conformer_ctc/jit_pretrained_decode_with_HLG.py \
+      --nn-model ./conformer_ctc/exp/cpu_jit.pt \
+      --HLG ./data/lang_char/HLG.fst \
+      --words ./data/lang_char/words.txt \
+      ./BAC009S0764W0121.wav \
+      ./BAC009S0764W0122.wav \
+      ./BAC009S0764W0123.wav
+
 Note that to generate ./conformer_ctc/exp/cpu_jit.pt,
 you can use ./export.py --jit 1
 """
@@ -23,12 +35,11 @@
 import math
 from typing import Dict, List
 
-import kaldi_hmm_gmm
 import kaldifeat
 import kaldifst
 import torch
 import torchaudio
-from kaldi_hmm_gmm import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
 from torch.nn.utils.rnn import pad_sequence
 
 

egs/yesno/ASR/tdnn/jit_pretrained_decode_with_H.py

@@ -28,7 +28,7 @@
 import kaldifst
 import torch
 import torchaudio
-from kaldi_hmm_gmm import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
 from torch.nn.utils.rnn import pad_sequence
 
 

egs/yesno/ASR/tdnn/jit_pretrained_decode_with_HL.py

@@ -28,7 +28,7 @@
 import kaldifst
 import torch
 import torchaudio
-from kaldi_hmm_gmm import DecodableCtc, FasterDecoder, FasterDecoderOptions
+from kaldi_decoder import DecodableCtc, FasterDecoder, FasterDecoderOptions
 from torch.nn.utils.rnn import pad_sequence
 
 

icefall/ctc/README.md

@@ -1,17 +1,17 @@
 # Introduction
 
 This folder uses [kaldifst][kaldifst] for graph construction
-and decoders from [kaldi-hmm-gmm][kaldi-hmm-gmm] for CTC decoding.
+and decoders from [kaldi-decoder][kaldi-decoder] for CTC decoding.
 
 It supports only `CPU`.
 
 You can use
 
 ```bash
-pip install kaldifst kaldi-hmm-gmm
+pip install kaldifst kaldi-decoder
 ```
 to install the dependencies.
 
-[kaldi-hmm-gmm]: https://github.com/csukuangfj/kaldi-hmm-gmm
+[kaldi-decoder]: https://github.com/i2-fsa/kaldi-decoder
 [kaldifst]: https://github.com/k2-fsa/kaldifst
 [k2]: https://github.com/k2-fsa/k2

requirements.txt

@@ -1,7 +1,7 @@
 kaldifst
 kaldilm
 kaldialign
-kaldi-hmm-gmm
+kaldi-decoder
 sentencepiece>=0.1.96
 tensorboard
 typeguard