Mysterious file
I can see in the documented templates that (from the image of the seamless expressive architecture), the pretssel template is directly completed with the vocoder (pretssel_mehifigan_vm.pt), and that's fine by me.
What I don't understand is that there's a file on hugginface called pretssel.pt (https://huggingface.co/facebook/seamless-expressive/blob/main/pretssel.pt), I understand that it's the pretssel model (described in the architecture), but it's heavier (802MB) than pretssel_melhifigan_wm (377MB )? (Pretssel + vocoder Mel hifigan < pretssel ??!!)
I was wondering if we could further increase the quality of the generations, so if we could use this file I'd be interested in how we could interfere with it (the “pretssel” card doesn't exist on the GitHub documentation).
thanks to all who reply :)