Hifigan Vocoder #16
Hifigan Vocoder #16
Conversation
Fix URL of MelGAN paper
Update contribution guidelines
2) Add remove weight norm layer on Generator
2) Rename MPD disc 3) Re-format remove weight norm generator
* Update code and integrate training code
|
Any updates @rishikksh20 ? |
|
I'm interested in HifiGAN too. Excited to do a training on my "Thorsten" dataset. |
|
@erogol Yes I have trained model on different configuration but not able to train successfully, I even copy paste original hifi-gan with exact settings and only use TTS data loader with pre-processing and yet result is same the audio is not understandable at all I think I need to go through TTS pre-processing thoroughly because I think I miss something there, also I have implemented l1 loss mentioned in hifigan paper and training on original hifigan architecture lets see how it's turnout. Also original hifigan takes a week to get trained so it takes time. |
|
@rishikksh20 do you have a certain timeline on this? I try to plan the next release and see if you are up for it. Feel free to tell if time is scarse, just asking for planing. |
|
@erogol it will finished within end of first week of April. |
|
@erogol One more think when I replaced TTS melspectrogram pre-processing with mine melspectrogram calculation it's working fine I think the problem is with pre-processing only. |
|
I guess we can retire this as we merged HiFiGAN. Thanks, @rishikksh20 for initiating the work. |
ci: combine coverage reports and merge test workflows
No description provided.