Orpheus TTS

Canopy Labs built Orpheus TTS on a Llama-3B backbone, treating speech generation as a language modeling problem, and the result became one of the most widely adopted open text to speech models of 2025. The English 3B model ships in finetuned and pretrained variants, the latter trained on over 100k hours of English speech, with zero-shot voice cloning, guided emotion through inline tags such as laugh and sigh, eight named voices, and streaming inference at roughly 200 ms latency, reducible to about 100 ms with input streaming. A multilingual research preview covers seven language pairs, and the repository includes a finetuning toolkit and audio watermarking support, with 1B, 400M, and 150M variants on the roadmap. Installation is a single pip install orpheus-speech, which pulls in vLLM for realtime GPU serving; community GGUF conversions allow slower CPU inference through llama.cpp. Code and weights are Apache-2.0, so commercial use is permitted, and the repo has grown past 6,000 GitHub stars. Developers embed it in voice agents and realtime assistants that need a self-hostable alternative to hosted speech APIs.