The reconstructed speech or audio is strange

[sounddemos]

I use the inference code and the weight.pth for analysis-synthesis, the speech can not be generated normally.

1.wav

[sounddemos]

1_orginal.wav is the groundtruth input

[bigpon]

Did you use H100? Try to set cudnn to false or use cpu for inference. The model has a known issue with H100 when cudnn is enabled.

[sounddemos]

Yes, I used H20 for inference. Now I use cpu for inference, the model can work well