Feature: Support for some non-verbal tags

[coezbek]

Not an issue, just a short note: I have played in the last two weeks with PlayDiffusion to test if it can be fine-tuned for non-verbal tags (<laugh>, etc.) and have released a fine-tune of the PlayDiffusion model which support three tags: <laugh>, <breath> and <pause>.

I upgraded the tokenizer to support the tags, add syllable estimation for the tags and trained the embeddings for the MaskGCT model to support the new tags. It certainly isn't perfect, because the amount of data for this available isn't that much, but it is working quite okay for many voices.

You can run the gradio directly from uvx:

uvx --python 3.12 playdiffusion@git+https://github.com/coezbek/PlayDiffusion

[nguyen-khang-ntq]

@coezbek Can you public the training code? I want to finetune this model in my language. Thank you.

[platform-kit]

@coezbek seconding request for training code!