Slue PR configs

[siddhu001]

Upload configs and provide access to best pre-trained models for SLT paper (https://arxiv.org/abs/2211.05869).

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (7c140c2) 76.54% compared to head (273c36f) 76.57%.
Report is 64 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #5087      +/-   ##
==========================================
+ Coverage   76.54%   76.57%   +0.03%     
==========================================
  Files         720      720              
  Lines       66639    66695      +56     
==========================================
+ Hits        51008    51072      +64     
+ Misses      15631    15623       -8     

Flag	Coverage Δ
test_configuration_espnet2	∅ <ø> (∅)
test_integration_espnet1	62.92% <ø> (ø)
test_integration_espnet2	50.10% <ø> (+0.02%)	⬆️
test_python_espnet1	19.06% <ø> (-0.02%)	⬇️
test_python_espnet2	52.45% <ø> (+0.05%)	⬆️
test_utils	22.15% <ø> (ø)

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

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[siddhu001]

Hi @ftshijt,

This PR is ready for review. The primary modifications include the addition of configuration files for the best-trained model in the SLT paper and the ESPnet-SLUE toolkit, which was presented at the ASRU workshop. I've also implemented some minor fixes to the scoring function and the 2-pass SLU implementation. Could you please review it?

[siddhu001]

Hi @sw005320,

This PR is related to slue-voxceleb and slue-voxpopuli recipes for ESPnet-SLUE toolkit. Please let me know if you have any comments before we can merge this PR!

[sw005320]

OK, I'll start to review.
@ftshijt, can you also start it?

[ftshijt]

Thanks for the update! Looks very good to me. Two minor comments

[ftshijt]

Consider make it standardize by setting conf/train.yaml

[siddhu001]

Done!

[ftshijt]

It seems to me that the change of many configs are quite similar, do we want to keep all of them? (I'm ok with it if that's for public benchmark, but ideally, it would be better if we could summarize them into fewer configs, which is also easier for users)

[siddhu001]

Yes, I considered it, but similar configs (like the ones using pre-trained ASR) have different sets of hyperparameters (like warmup steps, learning rate) selected based on validation performance. Hence, grouping similar configs would reduce reproducibility, so I think it is better to keep all of the configs.

[siddhu001]

@sw005320 I think this PR is ready to be merged! Let me know if you have any comments.

[sw005320]

The table is not correctly displayed.
Can you confirm it?

[siddhu001]

Thanks for catching this! I have fixed this now!