ASR2 recipe on Tedlium3 dataset

[kohei0209]

tedlium3/asr2 recipe

Implementation of tedlium3/asr2 recipe

• check data preparation stage (~ stage 8)
• check training stage (stage 8~)
• Get some results

[codecov]

Codecov Report

Merging #5331 (9f54c74) into master (5d0758e) will increase coverage by 2.64%.
Report is 485 commits behind head on master.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #5331      +/-   ##
==========================================
+ Coverage   72.72%   75.36%   +2.64%     
==========================================
  Files         679      709      +30     
  Lines       61692    65290    +3598     
==========================================
+ Hits        44865    49206    +4341     
+ Misses      16827    16084     -743     

Flag	Coverage Δ
test_configuration_espnet2	∅ <ø> (∅)
test_integration_espnet1	65.67% <ø> (-0.06%)	⬇️
test_integration_espnet2	48.71% <ø> (?)
test_python_espnet1	19.16% <ø> (-1.11%)	⬇️
test_python_espnet2	51.39% <ø> (-0.71%)	⬇️
test_utils	23.10% <ø> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

see 120 files with indirect coverage changes

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[simpleoier]

Please update the time information later.

[simpleoier]

Is the time information correct?

[simpleoier]

Ditto.

[sw005320]

What is the status of this PR? We can make it from a draft to regular PR if it is ready.

[simpleoier]

Once @kohei0209 gets the asr2 results with new config, we can bring this PR to regular and proceed to merge.

[simpleoier]

Hi @kohei0209 , can you continue this PR and upload you checkpoints?

[kohei0209]

I am sorry for the late reply. I'll upload the checkpoints and update this PR

[kohei0209]

Hi @simpleoier, is it okay to include the data filtering process for removing empty text at stage 6 in this PR? Since Ted3 has some empty texts, data filtering is necessary in asr2.sh.
If okay, could you tell me a good way to do so? I did data filtering as follows but it's dirty.

# remove empty text
cat "${data_feats}/org/${dset}/text.ts.en" | awk ' { if( NF != 1 ) print $0; } ' > "${data_feats}/${dset}/text.ts.en"
# align keys
# maybe fix_data_dir.sh should be used, it's dirty
utils/filter_scp.pl "${data_feats}/${dset}/text.ts.en" "${data_feats}/org/${dset}/utt2spk" > "${data_feats}/${dset}/utt2spk"
utils/filter_scp.pl "${data_feats}/${dset}/text.ts.en" "${data_feats}/org/${dset}/text.rm.${kmeans_feature_type}_${layer}_km${nclusters}" > "${data_feats}/${dset}/text.rm.${kmeans_feature_type}_${layer}_km${nclusters}"
utils/utt2spk_to_spk2utt.pl "${data_feats}/${dset}/utt2spk" > "${data_feats}/${dset}/spk2utt"

[sw005320]

Did you observe some improvements with it?
I'm curious about it.

Yes, stage 6 is the correct place.
But as you said, fix_data_dir.sh might be better.
Please also display how many utterances are pruned.
fix_data_dir.sh includes such information.

[kohei0209]

Thank you for your answer. I'll try fix_data_dir.sh.
This data filtering is for removing empty text (or audio without any utterance) and avoiding errors in the training stage, not for improving performance. I got an error in the training stage when I did not remove empty text.

[sw005320]

I see.
Theoretically, it should not have errors, but maybe there are some issues.

You do not have to do it, but one approach would be to add a special silence token for such utterances.

[kohei0209]

I am very sorry for the late reply.
I re-run the recipe without removing the data with empty reference text and got the following error that said the text.ts.en file does not include a space between the first and the second columns. This happens because the row with an empty reference text only has the utterance ID.

Original Traceback (most recent call last):
File "/mnt/aoni04/saijo/hackathon-2023summer/espnet-2023-720/tools/miniconda/envs/espnet_hackathon/lib/python3.10/site-packages/torch/utils/data/_utils/worker.py", line 302, in _worker_loop
data = fetcher.fetch(index)
File "/mnt/aoni04/saijo/hackathon-2023summer/espnet-2023-720/tools/miniconda/envs/espnet_hackathon/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 34, in fetch
data.append(next(self.dataset_iter))
File "/mnt/aoni04/saijo/hackathon-2023summer/espnet-2023-720/espnet2/train/iterable_dataset.py", line 176, in iter
raise RuntimeError(
RuntimeError: This line doesn't include a space: <_io.TextIOWrapper name='dump/raw/train_sp/text.ts.en' mode='r' encoding='utf-8'>:L39787: ApolloRobbins_2013G-0042011-0042121
)

I tried to use fix_data_dir.sh but I needed to filter utt2spk first because fix_data_dir.sh does not seem to refer to the text.ts.en file. But if we do so we cannot know how many utterances are pruned since utt2spk is already filtered. Are there any good solutions...?

cat "${data_feats}/org/${dset}/text.ts.en" | awk ' { if( NF != 1 ) print $0; } ' > "${data_feats}/${dset}/text.ts.en"
# need to filter utt2spk first because fix_data_dir.sh does not seem to refer to text.ts.en
utils/filter_scp.pl "${data_feats}/${dset}/text.ts.en" "${data_feats}/org/${dset}/utt2spk" > "${data_feats}/${dset}/utt2spk"
# filter the other files (spk2utt and text.rm.wavlm_large_21_km*)
utils/fix_data_dir.sh --utt_extra_files "${utt_extra_files}" "${data_feats}/${dset}"

[sw005320]

I see.
Then, I think it is fine with this current way.
But, how about adding lines to echo the number of sentences?
This is just for monitoring purposes and not so strict.

[kohei0209]

Thank you for your advice. I've added the code to show how many samples are removed:

# Remove empty text
cat "${data_feats}/org/${dset}/text.${tgt_case}.${tgt_lang}" | awk ' { if( NF != 1 ) print $0; } ' > "${data_feats}/${dset}/text.${tgt_case}.${tgt_lang}"
utils/filter_scp.pl "${data_feats}/${dset}/text.${tgt_case}.${tgt_lang}" "${data_feats}/org/${dset}/utt2spk" > "${data_feats}/${dset}/utt2spk"
utils/fix_data_dir.sh  --utt_extra_files "${utt_extra_files}" "${data_feats}/${dset}"
# check how many samples are removed
org_num_samples=$(wc -l "${data_feats}/org/${dset}/utt2spk" | cut -d' ' -f1)
filtered_num_samples=$(wc -l "${data_feats}/${dset}/utt2spk" | cut -d' ' -f1)
echo "filter samples with empty texts: removed $((org_num_samples - filtered_num_samples)) samples with empty text"

The log is as follows:

2023-10-14T22:06:53 (asr2.sh:316:main) Info: The valid_set 'dev' is included in the test_sets. '--eval_valid_set true' is set and 'dev' is removed from the test_sets
2023-10-14T22:06:53 (asr2.sh:583:main) Skipped stages:  11 16 17 18 
2023-10-14T22:07:02 (asr2.sh:813:main) Stage 6: Data filtering: dump/raw/org -> dump/raw
dump/raw/train_sp
fix_data_dir.sh: kept all 804780 utterances.
fix_data_dir.sh: old files are kept in dump/raw/train_sp/.backup
filter samples with empty texts: removed 9 samples with empty text
dump/raw/dev
fix_data_dir.sh: kept all 507 utterances.
fix_data_dir.sh: old files are kept in dump/raw/dev/.backup
filter samples with empty texts: removed 0 samples with empty text
2023-10-14T22:10:08 (asr2.sh:1697:main) Successfully finished. [elapsed=217s]

[kohei0209]

@simpleoier BTW, do you plan to switch the input orders of src and tgt texts (tgt first -> src first) in stage 13? (As we discussed before, it can make training more memory-efficient and faster)

[simpleoier]

@kohei0209 Thanks for the reminder. You can adjust the order in this PR.

[kohei0209]

I reflected your comments. I've also uploaded the model parameters on huggingface.
Could you make this PR from a draft to regular PR and check the changes?

[sw005320]

Thanks, @kohei0209!