mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-08-09 10:02:22 +00:00
Update AIShell recipe result (#140)
* add MMI to AIShell * fix MMI decode graph * export model * typo * fix code style * typo * fix data prepare to just use train text by uid * use a faster way to get the intersection of train and aishell_transcript_v0.8.txt * update AIShell result * update * typo
This commit is contained in:
pingfengluo
GitHub
parent
89b84208aa
commit
d1adc25338
@ -1,22 +1,22 @@
|
||||
## Results
|
||||
|
||||
### Aishell training results (Conformer-MMI)
|
||||
#### 2021-12-01
|
||||
(Pingfeng Luo): Result of <https://github.com/k2-fsa/icefall/pull/137>
|
||||
#### 2021-12-04
|
||||
(Pingfeng Luo): Result of <https://github.com/k2-fsa/icefall/pull/140>
|
||||
|
||||
The tensorboard log for training is available at <https://tensorboard.dev/experiment/dyp3vWE9RE6SkqBAgLJjUw/>
|
||||
The tensorboard log for training is available at <https://tensorboard.dev/experiment/PSRYVbptRGynqpPRSykp1g>
|
||||
|
||||
And pretrained model is available at <https://huggingface.co/pfluo/icefall_aishell_model>
|
||||
And pretrained model is available at <https://huggingface.co/pfluo/icefall_aishell_mmi_model>
|
||||
|
||||
The best decoding results (CER) are listed below, we got this results by averaging models from epoch 20 to 49, and using `attention-decoder` decoder with num_paths equals to 100.
|
||||
The best decoding results (CER) are listed below, we got this results by averaging models from epoch 61 to 85, and using `attention-decoder` decoder with num_paths equals to 100.
|
||||
|
||||
||test|
|
||||
|--|--|
|
||||
|CER| 5.12% |
|
||||
|CER| 4.94% |
|
||||
|
||||
||lm_scale|attention_scale|
|
||||
|--|--|--|
|
||||
|test|1.5|0.5|
|
||||
|test|1.1|0.3|
|
||||
|
||||
You can use the following commands to reproduce our results:
|
||||
|
||||
@ -31,12 +31,12 @@ export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7,8"
|
||||
python conformer_mmi/train.py --bucketing-sampler True \
|
||||
--max-duration 200 \
|
||||
--start-epoch 0 \
|
||||
--num-epochs 50 \
|
||||
--num-epochs 90 \
|
||||
--world-size 8
|
||||
|
||||
python conformer_mmi/decode.py --nbest-scale 0.5 \
|
||||
--epoch 49 \
|
||||
--avg 20 \
|
||||
--epoch 85 \
|
||||
--avg 25 \
|
||||
--method attention-decoder \
|
||||
--max-duration 20 \
|
||||
--num-paths 100
|
||||
|
||||
@ -1,5 +1,6 @@
|
||||
#!/usr/bin/env python3
|
||||
# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, Fangjun Kuang)
|
||||
# Copyright 2021 Pingfeng Luo
|
||||
#
|
||||
# See ../../../../LICENSE for clarification regarding multiple authors
|
||||
#
|
||||
|
||||
@ -1,6 +1,7 @@
|
||||
#!/usr/bin/env python3
|
||||
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang,
|
||||
# Wei Kang)
|
||||
# Copyright 2021 Pingfeng Luo
|
||||
#
|
||||
# See ../../../../LICENSE for clarification regarding multiple authors
|
||||
#
|
||||
@ -80,7 +81,7 @@ def get_parser():
|
||||
parser.add_argument(
|
||||
"--num-epochs",
|
||||
type=int,
|
||||
default=50,
|
||||
default=90,
|
||||
help="Number of epochs to train.",
|
||||
)
|
||||
|
||||
|
||||
@ -111,63 +111,66 @@ if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
|
||||
./local/compute_fbank_musan.py
|
||||
fi
|
||||
|
||||
lang_phone_dir=data/lang_phone
|
||||
lang_char_dir=data/lang_char
|
||||
if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
|
||||
log "Stage 5: Prepare phone based lang"
|
||||
lang_dir=data/lang_phone
|
||||
mkdir -p $lang_dir
|
||||
mkdir -p $lang_phone_dir
|
||||
|
||||
(echo '!SIL SIL'; echo '<SPOKEN_NOISE> SPN'; echo '<UNK> SPN'; ) |
|
||||
cat - $dl_dir/aishell/resource_aishell/lexicon.txt |
|
||||
sort | uniq > $lang_dir/lexicon.txt
|
||||
sort | uniq > $lang_phone_dir/lexicon.txt
|
||||
|
||||
./local/generate_unique_lexicon.py --lang-dir $lang_dir
|
||||
./local/generate_unique_lexicon.py --lang-dir $lang_phone_dir
|
||||
|
||||
if [ ! -f data/lang_phone/L_disambig.pt ]; then
|
||||
if [ ! -f $lang_phone_dir/L_disambig.pt ]; then
|
||||
./local/prepare_lang.py
|
||||
fi
|
||||
|
||||
# Train a bigram P for MMI training
|
||||
if [ ! -f $lang_dir/transcript_words.txt ]; then
|
||||
if [ ! -f $lang_phone_dir/transcript_words.txt ]; then
|
||||
log "Generate data to train phone based bigram P"
|
||||
aishell_text=aishell/data_aishell/transcript/aishell_transcript_v0.8.txt
|
||||
cat ${dl_dir}/${aishell_text} | cut -d " " -f 2- > $lang_dir/transcript_words.txt
|
||||
aishell_text=$dl_dir/aishell/data_aishell/transcript/aishell_transcript_v0.8.txt
|
||||
aishell_train_uid=$dl_dir/aishell/data_aishell/transcript/aishell_train_uid
|
||||
find data/aishell/data_aishell/wav/train -name "*.wav" | sed 's/\.wav//g' | awk -F '/' '{print $NF}' > $aishell_train_uid
|
||||
awk 'NR==FNR{uid[$1]=$1} NR!=FNR{if($1 in uid) print $0}' $aishell_train_uid $aishell_text | cut -d " " -f 2- > $lang_phone_dir/transcript_words.txt
|
||||
fi
|
||||
|
||||
if [ ! -f $lang_dir/transcript_tokens.txt ]; then
|
||||
if [ ! -f $lang_phone_dir/transcript_tokens.txt ]; then
|
||||
./local/convert_transcript_words_to_tokens.py \
|
||||
--lexicon $lang_dir/uniq_lexicon.txt \
|
||||
--transcript $lang_dir/transcript_words.txt \
|
||||
--lexicon $lang_phone_dir/uniq_lexicon.txt \
|
||||
--transcript $lang_phone_dir/transcript_words.txt \
|
||||
--oov "<UNK>" \
|
||||
> $lang_dir/transcript_tokens.txt
|
||||
> $lang_phone_dir/transcript_tokens.txt
|
||||
fi
|
||||
|
||||
if [ ! -f $lang_dir/P.arpa ]; then
|
||||
if [ ! -f $lang_phone_dir/P.arpa ]; then
|
||||
./shared/make_kn_lm.py \
|
||||
-ngram-order 2 \
|
||||
-text $lang_dir/transcript_tokens.txt \
|
||||
-lm $lang_dir/P.arpa
|
||||
-text $lang_phone_dir/transcript_tokens.txt \
|
||||
-lm $lang_phone_dir/P.arpa
|
||||
fi
|
||||
|
||||
if [ ! -f $lang_dir/P.fst.txt ]; then
|
||||
if [ ! -f $lang_phone_dir/P.fst.txt ]; then
|
||||
python3 -m kaldilm \
|
||||
--read-symbol-table="$lang_dir/tokens.txt" \
|
||||
--read-symbol-table="$lang_phone_dir/tokens.txt" \
|
||||
--disambig-symbol='#0' \
|
||||
--max-order=2 \
|
||||
$lang_dir/P.arpa > $lang_dir/P.fst.txt
|
||||
$lang_phone_dir/P.arpa > $lang_phone_dir/P.fst.txt
|
||||
fi
|
||||
fi
|
||||
|
||||
if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
|
||||
log "Stage 6: Prepare char based lang"
|
||||
mkdir -p data/lang_char
|
||||
mkdir -p $lang_char_dir
|
||||
# We reuse words.txt from phone based lexicon
|
||||
# so that the two can share G.pt later.
|
||||
cp data/lang_phone/words.txt data/lang_char
|
||||
cp $lang_phone_dir/words.txt $lang_char_dir
|
||||
|
||||
cat $dl_dir/aishell/data_aishell/transcript/aishell_transcript_v0.8.txt |
|
||||
cut -d " " -f 2- | sed -e 's/[ \t\r\n]*//g' > data/lang_char/text
|
||||
cut -d " " -f 2- | sed -e 's/[ \t\r\n]*//g' > $lang_char_dir/text
|
||||
|
||||
if [ ! -f data/lang_char/L_disambig.pt ]; then
|
||||
if [ ! -f $lang_char_dir/L_disambig.pt ]; then
|
||||
./local/prepare_char.py
|
||||
fi
|
||||
fi
|
||||
@ -181,7 +184,7 @@ if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
|
||||
if [ ! -f data/lm/G_3_gram.fst.txt ]; then
|
||||
# It is used in building HLG
|
||||
python3 -m kaldilm \
|
||||
--read-symbol-table="data/lang_phone/words.txt" \
|
||||
--read-symbol-table="$lang_phone_dir/words.txt" \
|
||||
--disambig-symbol='#0' \
|
||||
--max-order=3 \
|
||||
$dl_dir/lm/3-gram.unpruned.arpa > data/lm/G_3_gram.fst.txt
|
||||
@ -190,6 +193,6 @@ fi
|
||||
|
||||
if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
|
||||
log "Stage 8: Compile HLG"
|
||||
./local/compile_hlg.py --lang-dir data/lang_phone
|
||||
./local/compile_hlg.py --lang-dir data/lang_char
|
||||
./local/compile_hlg.py --lang-dir $lang_phone_dir
|
||||
./local/compile_hlg.py --lang-dir $lang_char_dir
|
||||
fi
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user