LightningFastSpeech2

LightningFastSpeech2 copied to clipboard

This project has gotten quite complex, and good documentation is needed. I will go with MKDocs again.

MiniXC

Metrics (such as DTW distance, mae, etc) should a) be modular to allow adding new ones easily b) include wasserstein distance and others

MiniXC

There should be parameter groups (with a certain prefix, e.g. "fastdiff_vocoder"), as well as config files. It might be worth looking into using an existing project for this or creating...

MiniXC

Currently e.g. dataset caching is handled in training code, which it shouldn't be. It is worth considering separating the TTSDataset out into it's own respository, or at least have it...

MiniXC