POC - llasa-3b

[kamomehettapi]

Hi I do not know if your interested but I have made Proof of Concept of this recent Model:

https://huggingface.co/spaces/srinivasbilla/llasa-3b-tts
https://huggingface.co/HKUSTAudio/Llasa-3B

I'm sorry if code is incomplete really it shows you can select and generate with Llasa zero-shot Model

I got pretty good results with clean speech sample

But, environment needs update to CUDA 12.4 due to requirements

Also I have weird issue where transformers gives whole prompt plus generated, and I try truncating to only generated but it's not always worked

And also you must install "xcodec2" module in repo root for work, their version on PIP is strangely different and not compatible

[erew123]

Hi @kamomehettapi I am looking at updating Pytorch etc. There are other impacts so this may take me a little bit of time to resolve. Thanks for the PR though and assuming I can get Pytorch updated, I will pull this in.

Thanks