Orpheus-TTS

Why we included this project

Orpheus is an open-source text-to-speech engine built on a Llama-3b backbone, and the LLM foundation shows in the output: speech carries natural intonation and rhythm instead of the flat delivery typical of older engines. You can steer expressiveness with simple emotion tags, clone a voice from a short reference without fine-tuning, and stream audio at roughly 200ms of latency, which suits live voice assistants and interactive apps where waiting for a full render is not an option. The repo ships a production-finetuned model and a pretrained one trained on 100k+ hours of English speech, plus data-processing scripts and sample datasets if you want to fine-tune on your own voices. A multilingual research release and a Baseten integration for optimized fp8/fp16 inference round it out, so teams that want human-sounding speech have a viable path to running it themselves.