fishaudio

/

openaudio-s1-mini

like 536

Follow

Fish Audio 903

I have the model working well - but each chunk I am getting pretty different sounding TTS voices. I have played with passing extra parameters to try and lock it in:

for example:

{
"temperature": 0.4,
"top_p": 0.15,
"repetition_penalty": 1.1,
"chunk_length": 120,
"max_new_tokens": 512,
"seed": 42000
}

I thought maybe recording my own voice would help keep this expressive, yet also constrained to a voice profile. Is there a way to record my own voice for a no-shot or few-shot version of the model. also, if so, is there a guide for this? I don't see any obvious options for it on the S1-mini, am I missing something or does only the s1 model support the no-shot/few-shot features?

Thanks!