Rishikesh (ऋषिकेश)

@Miralan implements the Time domain losses hope she might gives good input on this, I haven't rechecked this code for while as I am busy with other model implementation. Otherwise...

@Miralan yes I think that not have much impact on overall quality, it just helps to removes minute artifacts and jitters.

Also @Miralan have you worked with Wavenet architecture ?

Cuda 11, pytorch 1.8.2

Use Torch 1.8.0 with Cuda 11.1

I am using Cuda 8 with Nvidia driver 384

@WelkinYang Have you figured it out, why samples are average before ?

@WelkinYang Exactly I am also working to replace whole Text Encoder with FS2 Encoder like Diff-TTS.

@tdplaza it's the predicting performance of the un-conditional Parallel wavenet.

Nope, tinyVITS also use similar architecture for Waveform decoder but this repo solely code iSTFTNet from this paper: https://arxiv.org/pdf/2203.02395.pdf