search

Mangio621 / Mangio-RVC-Fork

*CREPE+HYBRID TRAINING* A very experimental fork of the Retrieval-based-Voice-Conversion-WebUI repo that incorporates a variety of other f0 methods, along with a hybrid f0 nanmedian method.
MIT License
996 stars 215 forks source link

[Help me!] All Training attempts generated the same voice. #204

Closed Ahm3dRN closed 7 months ago

Ahm3dRN commented 7 months ago

I've been trying to train my own model on a dataset I gathered. I used more than one person's voice, I first tried creating a model of a voice the dataset was 12 minutes of audio with 50 epochs the voice was nothing near. then I tried 100 epochs same exact voice. I changed the dataset to a different person as a different model and it's the same exact voice again. I tried a different dataset I used this time for 5x 2 min audio files. all the audio files are of good quality. I did this with 5 different models 5 different datasets (voices) and ranging epochs of 50-200 and they all generated the same exact voice regardless. I'm really new to this I need help to figure out what's wrong.

I'm running on NVIDIA GeForce GTX 1660 SUPER

Ahm3dRN commented 7 months ago

Did you manage to solve this problem?

Yes I just downloaded and installed a stable version from the hugging face repo. I ended up using the main repo.

Dubita commented 4 months ago

Thank you so much for helping! I was getting crazy all day trying to fix this

Edit: After 2 succesfully trained models, the third model got the same voice as before the main repo. I will try to test a few more things