Open nathanodle opened 1 year ago
Apparently there weren't any modifications (see README), so I'd use their repo for training. I don't understand how their datasets work as they seem to depend on a library, which is honestly pretty annoying
Edit: looking at another issue, seems there's language embeddings? Honestly an easy training script implementation would be nice
https://github.com/bensonbs/VALL-E-X-Trainer
this fork fixes a few key issues and supports a slightly altered ljspeech format
@korakoe Why in the https://github.com/0417keito/VALL-E-X-Trainer-by-CustomData/blob/8ff6a7987b46f72b8a8d8cabbd71979f39318f80/models/vallex.py#L826, you don't add the language embedding to the text like the inference
did
I'd like to retrain the checkpoint with a larger dataset but not sure what your training script modifications are compared to the lifeiteng repo. Would you be up for discussing sometime on Discord?