Pull request johan BUT sre

[gulamungon]

SRE recipe using CTS superset + voxceleb as embedding extractor training data (See README). There are very few changes outside the recipe. Let me know if this is not appropriate:

• In tools/make_shard_list.py changed so that if VAD info does not exist for a recording this recording will be used as is, i.e. no VAD will be applied. This is not ideal since absence of VAD info currently occurs also if VAD was estimated but no speech was found. However, there may be situations where we don't want to apply any sets, e.g., we may want to apply VAD to CTS but not voxceleb. Then we need it this way. This means that utterances for which VAD was ran but no speech detected, should be filtered before the shards are created. This will be the case if the sets are filtered with local/filter_utt_accd_dur.py since this script discards recordings with no Speech according to VAD. Ideally this should be improved so that recording for which for which no speech was detected will be removed while file that we don't want to apply VAD to will be kepts. Possibly by
  • Change the VAD info format so that it also contains files with no speech, then discard them in tools/make_shard_list.py while keeping the ones with no VAD info, i.e., those for which we do not want to apply VAD.
  • Keep the VAD format as is, i.e., no info for recordings with no speech but instead add a fake VAD info for the files we don't want to apply VAD for. This info would simply mark the whole segment as speech.
• Changed in get_data_for_plda(... in plda_utils.py to give a warning instead of a crash in whenever an entry is in the scp but not in utt2spk. This embedding will be skipped. (This can be the case if files have been added added to the original CTS data folder since the scp is created by finding all wav files in this directory. In the case of BUT, we have some extra files here for sanity checks.) If this solution is not appropriate we can change in the data preparation script so that mismatch between the created scp and utt2spk is fixed already there.
• Changed in local/make_system_sad.py (this change is only in the local recipe) so that it process so that VAD is processed for a limited number of files at the time (hardwired to 10000), after which the VAD result is saved (Instead of processing all at once). When extracting VAD. It took very long time to start otherwise. This could also be helpful in case there is a crash since output is saved after each part instead of after the whole set.

[JiJiJiang]

Please fix Lint errors.

[czy97]

Hello Johan @gulamungon, thanks for the contribution. Can you first fix the Lint errors. Locally, you can use the flake8 command to check for problematic files and use yapf -i xxx.py to automatically format the problematic files.

[gulamungon]

Hello Johan @gulamungon, thanks for the contribution. Can you first fix the Lint errors. Locally, you can use the flake8 command to check for problematic files and use yapf -i xxx.py to automatically format the problematic files.

Sure, I'll try to fix it asap.

[gulamungon]

I fixed it hopefully.

[gulamungon]

Changed tabs to spaces.

[wsstriving]

@gulamungon Hi, Johan, thanks for the contribution, it seems there are still some lint errors (trailing whitespaces) @czy97 @JiJiJiang Maybe you guys start the reviewing first, and we do the lint fix afterwards

[gulamungon]

Hi Shuai,

Ok. I see. I can also try to fix it but in the coming two weeks I'm quite busy so most likely it will not manage during this time.

Best, Johan

On Sun, Jun 16, 2024 at 1:59 AM Shuai Wang @.***> wrote:

@gulamungon https://github.com/gulamungon Hi, Johan, thanks for the contribution, it seems there are still some lint errors (trailing whitespaces) @czy97 https://github.com/czy97 @JiJiJiang https://github.com/JiJiJiang Maybe you guys start the reviewing first, and we do the lint fix afterwards

— Reply to this email directly, view it on GitHub https://github.com/wenet-e2e/wespeaker/pull/326#issuecomment-2170978739, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAXXG3A2DZY6V5LQLPAASBTZHTIOXAVCNFSM6AAAAABIYYC7OKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCNZQHE3TQNZTHE . You are receiving this because you were mentioned.Message ID: @.***>