Thoughts on auxiliary audio losses using V-Diffusion

[brentspell]

First, thanks for creating this repo, it is a great resource for audio ML.

The Moûsai paper hints at additional/perceptual losses in the Future Work section. I'm curious whether this would be possible to do in the V-Diffusion framework, since the denoiser predicts the "velocity" of the noise instead of the clean audio. Do you know of a transformation that could be applied to the model outputs at training time, for comparing against ground truth using an additional criterion?