Hubert-discrete symbol-based HiFiGAN with duration predictor

[ftshijt]

• [x] major functions/code necessary for training
• [x] recipe for cvss_c
• [x] inference scripts for cvss_c with duration predictor

(I also apply isort to follow the style in ESPnet, please let me know if this is not needed for some reason)

[kan-bayashi]

Thanks, @ftshijt

(I also apply isort to follow the style in ESPnet, please let me know if this is not needed for some reason)

To make this PR's contribution clear, I will make another PR to apply isort. Please wait for a while.

[kan-bayashi]

I applied isort to master. Please merge it.

[ftshijt]

Hi @kan-bayashi , may I ask for your help on the CI test? I'm not quite sure why the CI failed for the import as I'm basically following the same way as other losses.

parallel_wavegan/losses/__init__.py:2:1: F403 'from .duration_prediction_loss import *' used; unable to detect undefined names
parallel_wavegan/losses/__init__.py:2:1: F401 '.duration_prediction_loss.*' imported but unused

[kan-bayashi]

I'm not quite sure why

Maybe due to simple typo lol

[ftshijt]

I think this PR is ready. Please let me know if you have any suggestions or comments~ many thanks!

[ftshijt]

Maybe I can also put some pre-trained models into the readme?

[kan-bayashi]

Maybe I can also put some pre-trained models into the readme?

That is great. Please open a new PR when the model is ready.

[kan-bayashi]

If you have a time, adding a test for the hifigan with duration prediction is preferable. Anyway, I will merge this PR at first :) Thank you again, @ftshijt.