How can I train Laura to use my voice for this model, so that it doesn't clone my voice but generates it directly using my own?

[persey01]

How can I train Laura to use my voice for this model, so that it doesn't clone my voice but generates it directly using my own?

[rodrigomatta]

I’m not 100% sure if I’m following you correctly, but if you’re looking to bake your voice directly into the model (native generation) rather than using zero-shot cloning, that’s not something s2.cpp supports. This repo is strictly for inference.

Beyond that, the official Fish Audio documentation actually strongly advises against fine-tuning the S2 Pro model. Since it was trained using Reinforcement Learning (GRPO), performing a standard fine-tune can shift the model's distribution and lead to significant performance degradation. This is likely why their finetuning guide only covers the S1-mini and doesn't even mention S2.

So, if your goal is consistency, 'training' a custom S2 model isn't the way to go right now. Your best bet is still focusing on high-quality reference audio, clean, with a matching transcript and the specific style you’re after to get the best out of the zero-shot cloning.

[mediaizy]

The problem is that it only works OK in very few languages , the Global Coverage languages have only base knowledge and for sure they need to be finetuned, not for a specific voice (because the cloning works surprisingly well) but for general knowledge of the specific languages.
I am also looking for the best way to do a finetune for a specific language but i cant find any documentation and they are not even answering on their github page.