Voice changing randomly

[taresh18]

Backend impacted

The Rust implementation

Operating system

Linux

Hardware

GPU with CUDA

Description

Hi, I am running the rust server for TTS. Using one of the custom voices available in the voice library. Its a male voice, however randomly across generations, it changes to a female voice. And for the next generation, it reverts back to male. Don't know what's the isssue here. Any idea how to fix this?

Extra information

None

Environment

Fill in the following information on your system.

• Operating system version:

If the backend impacted is PyTorch:

• Python version:
• PyTorch version:
• CUDA version (run python -c 'import torch; print(torch.version.cuda)'):
• GPU model and memory:

If the backend is MLX:

• Mac model:

[LaurentMazare]

We indeed noticed that this happens from time to time on specific voices but overall it should not be that common. Is it on a specific voice? Maybe you can try other voices to see if it also happens with them (though through our experience with unmute it should not be the case with the voices used there).

[taresh18]

@LaurentMazare yes, I am using this voice voice-donations/dwp.wav and its pretty common with it (3-4/10 times). Is this an issue with the model itself or something with the rust server? Any insights on how I can fix it?

[LaurentMazare]

Just tried it and indeed the voice is not preserved properly. I doubt that it's an actual bug in the model implementation and my guess is that it has more to do with the recording itself. If it's your voice donation, maybe you can create a few more entries there with different intonations and hopefully they won't show the issue this time.