tts : add support for Orpheus

[ggerganov]

HF: https://huggingface.co/collections/canopylabs/orpheus-tts-67d9ea3f6c05a941c06ad9d2

These TTS models seem suitable for supporting. To do that, we need to implement the SNAC audio codec: https://github.com/hubertsiuzdak/snac/

Sample implementation using Python-based inference of SNAC: https://github.com/isaiahbjork/orpheus-tts-local

Similar model support (OuteTTS): #10784
Can be used as a reference how to implement this.

[leoflowers]

Howdy! I'd like to give this issue a try.

[LostRuins]

It would be awesome if we could get xcodec, we already have plenty of tts but no ttmusic yet. Everything is a llama model and snac/wavtokenizer/xcodec the vocoding is all that's missing #11467

[ggerganov]

Everything is a llama model and snac/wavtokenizer/xcodec the vocoding is all that's missing

Technically yes, though we have to improve the way that these codecs are implemented and supported in general. Probably be able to have a single GGUF file with both the LLM and the codec, instead of separate. And be able to create either separate decoder / vocoder contexts from a single model. Or alternatively, a combined decoder+vocoder context. At least that's my general idea for supporting multi-modal cases, though I'm still figuring it out.

we already have plenty of tts

llama.cpp has only OuteTTS support via WavTokenizer. It would be nice to have at least one more TTS, so we can find some common patterns which would help to implement the above.

the vocoding is all that's missing

Mostly yes, but we also have to figure out how to do audio streaming. I am not sure yet how it works, but with Orpheus we should be able to understand it because it supports real-time streaming.

[scalar27]

This works pretty well on my Mac M1 running two instances of llama-server and fastrtc.
https://github.com/PkmX/orpheus-chat-webui