Add svoice recipe for wsj0-2mix speech separation

[nateanl]

• Add svoice separator proposed in Single Channel Voice Separation for Unknown Number of Speakers Under Reverberant and Noisy Settings
• Add multi-layer PIT wrapper that weighted sum the SI-SNR loss of multiple RNN layers' outputs.
• Add configuration file for wsj0-2mix experiment with svoice.

[sw005320]

• Can you merge the latest master?
• Can you add the result and model link to https://github.com/espnet/espnet/blob/master/egs2/wsj0_2mix/enh1/README.md?

[nateanl]

Sure. The model is under training, I will update the result table and upload the model once the training is finished.

[nateanl]

I uploaded the model to https://huggingface.co/Zhaoheng/svoice_wsj0_2mix

[sw005320]

@Emrys365, can you review this PR?

@nateanl, can you add a test?

[Emrys365]

Sure.

[Emrys365]

LGTM! I just made some minor comments.

[Emrys365]

Could you format the python scripts to pass black and flake8 style checks?

Also, could you add some unit tests in test/espnet2/enh/test_espnet_model.py and test/espnet2/enh/separator/test_svoice_separator.py?

You could take existing unit tests under test/espnet2/enh/ as a reference.

[codecov]

Codecov Report

Merging #4257 (9333684) into master (ff7c051) will increase coverage by 0.07%.
The diff coverage is 98.70%.

@@            Coverage Diff             @@
##           master    #4257      +/-   ##
==========================================
+ Coverage   80.71%   80.79%   +0.07%     
==========================================
  Files         453      456       +3     
  Lines       39576    39739     +163     
==========================================
+ Hits        31945    32107     +162     
- Misses       7631     7632       +1     

Flag	Coverage Δ
test_integration_espnet1	67.13% <ø> (ø)
test_integration_espnet2	49.87% <25.97%> (-0.19%)	⬇️
test_python	67.29% <98.70%> (+0.12%)	⬆️
test_utils	24.45% <ø> (ø)

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

Impacted Files	Coverage Δ
espnet2/enh/layers/dpmulcat.py	96.42% <96.42%> (ø)
espnet2/enh/loss/wrappers/multilayer_pit_solver.py	100.00% <100.00%> (ø)
espnet2/enh/separator/svoice_separator.py	100.00% <100.00%> (ø)
espnet2/tasks/enh.py	99.12% <100.00%> (+0.02%)	⬆️
espnet2/enh/loss/criterions/time_domain.py	92.98% <0.00%> (-3.10%)	⬇️
espnet2/st/espnet_model.py	0.00% <0.00%> (ø)
espnet2/bin/enh_scoring.py	94.20% <0.00%> (+1.44%)	⬆️
espnet2/enh/encoder/null_encoder.py	90.00% <0.00%> (+20.00%)	⬆️
espnet2/enh/decoder/null_decoder.py	100.00% <0.00%> (+28.57%)	⬆️

📣 Codecov can now indicate which changes are the most critical in Pull Requests. Learn more

[sw005320]

Thanks, @nateanl!