이준혁(Junhyeok Lee)
이준혁(Junhyeok Lee)
After this commit & Before this commit 
Since there are differences between distributions of the real and generated spectrogram, we do not recommend using 1.0.
Hello! I am not sure about your modified module, but it seems that it is caused by your forward arg speakers while you are using single speaker dataset https://github.com/jaywalnut310/vits/blob/2e561ba58618d021b5b8323d3765880f7e0ecfdb/models.py#L456 Check...
We did not try cross-lingual, but it seems to work with a cross-lingual setup. To train a cross-lingual Assem-VC, a multilingual Cotatron is needed. Since training Cotatron is almost identical...
Sorry for the late response. Due to the company's name change, it changed to [https://maum-ai.github.io/assem-vc/](https://maum-ai.github.io/assem-vc/).
Hi, it is a code for the checkpoint prefix and we fix that part with git commit hash. It seems that you do not run your code on git folder....
It seems that one of them is in cpu. Try to check PhaseAug is on cuda.
It does not contain any trainable variable, so identical to frozen. So it is identical results even if `train()` or `eval()`. I added it for clarity.
Hi sorry for late response. Since MRD only uses magnitude of STFT, PhaseAug could not be adopted to MRD. PhaseAug is only designed for waveform discriminators. Since BigVGAN also utilize...
Sorry, I just tested it on a higher version of torch, without compiling. I will try it in a month.