What change need make to compatible with Hi-fi GAN?

[faranaziz]

I mean pre processing the mels. https://github.com/jik876/hifi-gan/issues/61

[seungwonpark]

If you're using the Hi-fi GAN (as shown in Assem-VC), then you should have a look at both mel calculation code and the configuration:

• mel calculation script: https://github.com/mindslab-ai/cotatron/blob/master/modules/mel.py
• config: https://github.com/mindslab-ai/cotatron/blob/master/config/global/default.yaml

You'll need to change the first one to https://github.com/jik876/hifi-gan/blob/master/meldataset.py#L49-L72, and the second one to https://github.com/jik876/hifi-gan/blob/master/config_v3.json#L17-L27. Hope that helps.

[faranaziz]

Thanks so much. Is this how you trained the new upcoming model?

[seungwonpark]

No. We changed the mel calculation script and configuration to match the Hifi-GAN’s version.

[faranaziz]

I make suggested change and get artifcat like reported here: https://colab.research.google.com/gist/tulasiram58827/8a7660ab21ca3ea246141bbe4b0f87c0/hifigan_issue.ipynb

[Rashi2011]

I am trying to take mel from fastspeech2 and put that is hifiGAN to generate audio but all I am getting is noise . please suggest some Ideas.