search

voicepaw / so-vits-svc-fork

so-vits-svc fork with realtime support, improved interface and more features.
Other
8.77k stars 1.18k forks source link

Results are very noisy? (Mac M1 Max) #1067

Open Sarang0218 opened 9 months ago

Sarang0218 commented 9 months ago

I am trying to train a model on my macbook (M1 Max) and even after 30k steps, it still sounds very noisy. The loss seems to converge (very slowly), but it still sounds unbearable.

Original Post:

I have a MacBook M3 Pro which I wanted to use to train.

The problem I am facing is that when I run infer on the M3 trained model, the result is some unbearable noise, not even near to human voice. I was curious what would happen if I continued the training which I have started on Colab. It gets totally messed up. So it's giving really bad results when I start the training from scratch or when I try to continue a previous training.

Is there some setup or config I am missing to be able to get it working?

Originally posted by @nagyniki017 in https://github.com/voicepaw/so-vits-svc-fork/discussions/1046

Sarang0218 commented 9 months ago

I trained my model for 30k steps (1665 epochs)

34j commented 7 months ago

Is it simply because the device of the model is not changed to CPU when saved? If so, I think #1149 will help