Convert vocos checkpoint to bigvgan

[kmn1024]

Checks

• This template is only for question, not feature requests or bug reports.
• I have thoroughly reviewed the project documentation and read the related paper(s).
• I have searched for existing issues, including closed ones, no similar questions.
• I confirm that I am using English to submit this report in order to facilitate communication.

Question details

Is there a fast way to convert a F5 checkpoint that was trained using vocos mels to one that is compatible with bigvgan? For example, can we freeze most of the parameters, and only continue-train a small number of parameters? Or must we train all parameters again?

[ZhikangNiu]

But we also release BigVGAN F5 checkpoint in the huggingface ( Why we need to convert the checkpoint?

[kmn1024]

Because I have finetuned many models on top of the vocos checkpoint. Now I wonder if I need to redo all that if I want bigvgan.

[ZhikangNiu]

Got it, this is a great idea. I think you can first compare the differences between these two MelSpec implementation methods, then add a preprocessing network or postprocessing network, and freeze all the dit blocks. You can try this approach to see if the model can map the vocos features space to the bigvgan features space on its own. （Intuitively, this seems to be the simplest approach