do you have plan to release the base model?

yl4579 commented 2 months ago

Due diligence

[X] I have done my due diligence in trying to find the answer myself.

Topic

The paper

Question

In the paper, the audio generation model was trained on millions of hours of data and it seems this model is well capable of generating any voice and speaking styles, while the released models were only finetuned with a single speaker. Is there any plan to release the base model before finetuning? I think it will be very significant to have such a big model available for the TTS community, although I understand the ethical concerns and it may not be possible to release a model capable of high fidelity voice cloning.

LaurentMazare commented 1 month ago

The short answer is "not immediately". I agree with you that it would be very helpful for the community to release the whole model but also share the safety concerns that you mentioned so overall it's not entirely clear to us what the best way forward is. We have no plans on releasing the weights for the base model in the very near future but may well revisit if we believe it's a safe thing to do.

alkeryn commented 1 hour ago

@LaurentMazare Any news on this ?

i'm not quite sure it is that big of a safety risk knowing voice cloning is very easy to do now (ie F5 tts, openvoice) and can be coupled with a normal llm and stt. moshi would allow for more realtime interaction which would be cool for practical uses but nothing that'd be actually better for fraudulent uses.

either way, thanks for the good work and what has been released so far !

kyutai-labs / moshi