Retrieval-based-Voice-Conversion-WebUI icon indicating copy to clipboard operation
Retrieval-based-Voice-Conversion-WebUI copied to clipboard

RVC voice is only a shrilling bip / ring noise !

Open Docteur-RS opened this issue 8 months ago • 4 comments

Trained my model on 200 epoch with all the steps and got the .pth and the index file !

However when I try to use the RVC file either in the the convert webUI or in w-okada I only get a ring sound. Like a BIP...

I had no particular errors during training. Only the There appear to be 20 leaked semaphore objects to clean up at shutdown warning. But it is logged only at the end. I don't think it's that bad.

Am I the only one that gets a biiip as voice ??

Using:

  • Ubuntu
  • Nvidia CUDA 12.8
  • H100

I'm so close. Any ideas are welcome 😄

Docteur-RS avatar Apr 03 '25 22:04 Docteur-RS

Did the training on an RTX4090 and it worked !

So there seem to be some king of silent incompatibility with RVC and H100 or Ubuntu... Could also be the version of CUDA... I don't know. There were no error left to fix so it sould have worked on the H100 like it did on the 4090.

Docteur-RS avatar Apr 09 '25 07:04 Docteur-RS

Hey @Docteur-RS , just curious what was your training config like? is the fp_16 on there by default, what was the loss looks like?

rasenganai avatar Sep 09 '25 07:09 rasenganai

It's too old. I don't remember much from the config. Sry.

The loss was complete as it only outputed BIIIIIIIIP when I was speaking through w-okada. The exact sames steps on Windows worked perfectly though.

Are you also experiencing something similar?

Docteur-RS avatar Sep 09 '25 08:09 Docteur-RS

On H100, after few epochs my loss_gen shoots to nan.

giving only silent audios. If i dont use fp_16, it still shoots to high value giving only silent.

rasenganai avatar Sep 09 '25 10:09 rasenganai