hexgrad

/

Kokoro-82M

like 2.65k

I too would like to know. I've tried typing things like "uh" and wrapping them in asterisks as well (i.e. uh), but to no avail. An expanded or retrained version would be cool to have it accept a certain common syntax for these. Like wrapping filler sounds with double asterisks or supporting a movie script like format in parentheses. Do we need additional data for this?

I dont know. Maybe if they could fine-tune it on filler words, then i think it would help.

Yea, that would be helpful. A phonetic syntax override would also be useful for all edge cases. I.e. /ʌ/ or /ʌm/ too.

Phoneme override should be available with the next phonemizer: https://github.com/hexgrad/misaki but if good filler word usage is not present in training, inference results may be lacking.

Also see #69

@hexgrad , is the phoneme override being tracked separately as a feature update in the roadmap for Kokoro or will its incorporation into misaki auto add the feature to kokoro if the dependency is updated?

Also, thanks for linking to #69. After reading that, may I ask if you are in need of more training data with filler words?