how about the quality of this net

[hdmjdp]

Have you got good audio?

[rishikksh20]

yes better than hifi-gan with less training

[hdmjdp]

yes better than hifi-gan with less training

but in my experiments, the stftnet has larger machine noise in the audio, did you have it?

also in the training, there is "max value is tensor(1.1451, device='cuda:0', grad_fn=<MaxBackward1>) " note

[ease-zh]

also in the training, there is "max value is tensor(1.1451, device='cuda:0', grad_fn=) " note

Since no activation is applied to spec, only an exp to make it positive, its upper bound is unlimited, which causes an unlimited bound for the generated audio. Leaving the spec unbounded is not a good idea.