Add espnet2 TTS recipe on M-AILABS

[Takaaki-Saeki]

This PR adds a TTS recipe on M-AILABS.
TTS recipes on M-AILABS was supported in espnet1 but not supported in espnet2.

This recipe only uses a en_US single-speaker data.
I have uploaded a pretrained Tacotron2 model to HuggingFace.

I also prepared a script for multilingual data preparation: local/data_multilingual.sh to train a TTS model on the whole dataset.

[codecov]

Codecov Report

Merging #4701 (0418427) into master (e9d583b) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master    #4701   +/-   ##
=======================================
  Coverage   80.45%   80.45%           
=======================================
  Files         527      527           
  Lines       46215    46219    +4     
=======================================
+ Hits        37181    37185    +4     
  Misses       9034     9034           

Flag	Coverage Δ
test_integration_espnet1	66.37% <ø> (ø)
test_integration_espnet2	49.05% <0.00%> (-0.01%)	⬇️
test_python	68.66% <100.00%> (+<0.01%)	⬆️
test_utils	23.30% <ø> (ø)

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

Impacted Files	Coverage Δ
espnet2/text/phoneme_tokenizer.py	82.15% <100.00%> (+0.26%)	⬆️

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

[ftshijt]

Looks great! Some minor suggestion: given that the local/data.sh is automatically called in the TEMPLATE, how about combine data.sh and data_multilingual.sh but with an additional augment? In that way, we can change the config simply by change the --local_data_opts

In addition, could you also update the entry in egs2/TEMPLATE/README.md ?

[Takaaki-Saeki]

Thanks for the suggestion!
I have merged the en_us and multilingual recipes.
Also updated the readme.

[Takaaki-Saeki]

Hmm... it seems that some tests regarding espnet2/slu/test_transcript_espnet_model.py are not successful.

[sw005320]

Hmm... it seems that some tests regarding espnet2/slu/test_transcript_espnet_model.py are not successful.

@siddhu001, could you check this?
some interfaces are changed?

[siddhu001]

@sw005320 It seems to be problem with hugging face.

E       TypeError: register_buffer() got an unexpected keyword argument 'persistent'
tools/venv/lib/python3.7/site-packages/transformers/models/bert/modeling_bert.py:198: TypeError

I think it is because of some huggingface version update. I am looking into this. @Takaaki-Saeki if you have any suggestions, please let me know!

[Takaaki-Saeki]

Hi @siddhu001,

I think it is related to the version of PyTorch.
BertEmbeddings class requires persistent=false in nn.Module.register_buffer().
See this.

But this arg is not implemented in PyTorch v1.5.0 (see this), and it is introduced in v1.6.0 (see this).

So could you take a look at the pytorch version?

[ftshijt]

Many thanks! Merged.