fix: save processed audio in wav instead of pt

[MENGZHEGENG]

Solve #336
Now it saves the processed audio in wav instead of pt.

Related PR:

• feat: read audio from wavs instead of pt FastSpeech2_lightning#68
• feat: read wav instead of pt HiFiGAN_iSTFT_lightning#29

[MENGZHEGENG]

Hi, finally this PR is ready for review :)

Look forward to your feedback!

[codecov]

The author of this PR, MENGZHEGENG, is not an activated member of this organization on Codecov.
Please activate this user on Codecov to display this PR comment.
Coverage data is still being uploaded to Codecov.io for purposes of overall coverage calculations.
Please don't hesitate to email us at [email protected] with any questions.

[github-actions]

CLI load time: 0:00.30
Pull Request HEAD: 9b8ecf051333b41f86047372a0cee0663986cab0
Imports that take more than 0.1 s:
import time: self [us] | cumulative | imported package

[roedoejet]

Fantastic - looks great to me @MENGZHEGENG . Just some very minor changes requested in the submodules. Everything else looks good. I tested preprocessing out on my machine and it worked great. I didn't try training yet.

Could you also update the e2e dataset loader to use .wav and not .pt? https://github.com/roedoejet/EveryVoice/blob/03dca0e270e6cf173c12f8b71bd4139a9aa8db88/everyvoice/model/e2e/dataset.py#L102

[marctessier]

It looks good Tim.

I di a few extra tests to compared training times using this and the latest version of EV. using 1000 entries of LJ .. Both took ~ 1h2minutes. So it looks like it does not impact the training performance.

Save_wave 11:06 -> 12:08
stock 11:24 —> 12:26

[joanise]

Small detail: once your PRs are ready, you should rename them to remove the [WIP].

[MENGZHEGENG]

Thanks Eric for the reminder and I've removed the [WIP].

[roedoejet]

thanks so much Tim, this will be very helpful! great work

[joanise]

Cleaned up into two commits, looks great, thanks!

[MENGZHEGENG]

Thank you so much Eric for your guidances and instructions on cleaning the commits!