Open litsa-the-dancer opened 2 months ago
Thanks for your advice. I will start doing this after the code being refactored into a programmer-friendly style. Anyone who has interest can also help us by mentioning good algorithms or open PRs.
I don't think that the change to other gans will generate a substantial change in quality, the current quality of RVC is quite good, where I think there is potential is mainly in the optimization of the new architectures
I don't think that the change to other gans will generate a substantial change in quality
Yes, the hifigan in RVC is modified so it has already been bound in the algorithm.
where I think there is potential is mainly in the optimization of the new architectures
Yes, I will do this first. This issue will be considered lastly.
Applio developers are already implementing RVC v3 with BIGVGAN as an experiment in their repository:
https://github.com/IAHispano/Applio/tree/exp-bigvgan
I think we can contribute to that or merge in this repository the changes that they're doing. Idk if the devs here want to contribute with blaise-tk (developer of Applio).
@fumiama @alexlnkp
Applio developers are already implementing RVC v3 with BIGVGAN as an experiment in their repository:
https://github.com/IAHispano/Applio/tree/exp-bigvgan
I think we can contribute to that or merge in this repository the changes that they're doing. Idk if the devs here want to contribute with blaise-tk (developer of Applio).
@fumiama @alexlnkp
currently it is a bit broken
I will continue refacturing this repo first. After that, I will consider algorithm improvement.
Since hifi-gan is really old and outdated, i suggest you add something newer and more robust such as BIGvae gan. The results seem promising and can improve the overall performance of the model and potentially help with stuff such as mode collapse and noise handling. All in all, it's about time RVC gets a newer and more efficient GAN. ^w^