facebook

/

mms-tts

like 106

This is because the model was not trained on numerals (e.g. 123 or special characters (e.g. $), only lower-cased text. We can probably update the pre-processing in the tokenizer to map from any numerals/special characters to their text counterparts

cc @ylacombe (as this is what we recently did for SpeechT5 too)

We are very looking forward to the recent release to handle digital issues.

I have routines to turn numbers into words instead, that works okay.

https://github.com/groovybits/consciousChat/blob/main/characterChat.py#L45

I have tried to handle the multi-lingual cases, mostly working I think but haven't tested heavily.

It's really cool, but where are female voices? I am needing both male and female ones ideally.

Also is there a way to have different consistent seed'd character voices? Sorry if this is a naive question.