Special tokens to control pausing?
#3
by
mstachow
- opened
The model appears to randomly ignore periods or commas and the speech sounds a bit odd. Are there special pause characters or something to get it to stop and take a breath?
The model is trained on neither commas and periods, so they are filtered from the input text via a normalisation step in the pre-processing: https://github.com/huggingface/transformers/blob/910faa3e1f1c566b23a0318f78f5caf5bda8d3b2/src/transformers/models/vits/tokenization_vits.py#L127
Using hyphens is indeed the best option here
Will this model be re-trained on commas? even use hyphen, it still sounds very odd, just a bit better than commas. Besides, I found some pronounce is wrong, you can test "library".