Training guidline

[alealv]

Training guidline

[synesthesiam]

Hi @alealv, you're welcome 🙂

I've updated the repo to include some example config files and the README now has an example of running the training command.

1. This is correct; librosa will take care of resampling them as well
2. Correct. See the example configs. If you don't provide a config, the default settings will be used and one will be written on each checkpoint.
3. This was my original plan, but it seems the training involves computing mels with a different fmax for loss (None vs 8000 Hz). I haven't tested if removing this would change anything.
4. There's no speaker embedding, but the original authors trained on VCTK so there were definitely multiple speakers.
5. No validation steps, yeah; I was just manually checking the output from my checkpoints.
6. Caching the mels would probably help performance; you'd have to make sure the audio splitting code chops them up correctly, and there would need to be a separate cache for the loss mels (no fmax)
7. I don't, unfortunately. It seems the original authors ran their training for several million steps. I never made it that far from scratch.