voicebox-pytorch

Why we included this project

Voicebox is Meta's flow-matching text-to-speech model, notable for bringing in-context learning to speech generation, and this repository is a PyTorch reimplementation of it. The code wires the whole pipeline together, from a text-to-semantic module through an EnCodec-based audio codec to a conditional flow matching wrapper, so you can train on raw audio and generate speech conditioned on both text and a reference clip. It also supports Spear-TTS conditioning, which the author reports converges well in practice. One honest caveat: the maintainer now points people to E2 TTS as the more current option, so treat this as a solid reference for studying how Voicebox works and for experiments, not as a production-ready service.