interesting idea, does not really work #12
Description
While the attempt to use iSTFT to avoid upscaling artifacts introduced by last two upscales in HiFiGAN is commendable, it just does not work. On a single sample the model performs better than MRF HiFiGAN, but on a large training set this model falls short.
The upscaling artifacts from prior states are still there, and the Resblock applied to harmonic guidance fails to propagate them to higher frequencies so with just two upscaling resblocks is it not enough to get any meaningful resolution over 6KHz. Using 8 extra harmonics for the sinegen is the only thing that prevents it from failing completely.
