Kyutai TTS

Kyutai's delayed-streams-modeling repository carries the production releases of the technique behind Moshi: aligned audio and text streams are modeled framewise with a deliberate delay between them, which turns both speech-to-text and text-to-speech into streaming problems. The TTS models consume text as a stream, so audio playback can begin while an upstream LLM is still generating its reply, removing the usual wait for a complete sentence; the STT side offers a roughly 1B parameter English and French model with 0.5 second latency and semantic voice activity detection, plus a 2.6B English model. Kyutai ships three implementations: PyTorch for research, a Rust server with WebSocket streaming for production, where a single H100 can serve about 400 realtime streams, and MLX for Apple silicon down to iPhone class devices. Installation goes through the moshi Python package with pip or uv. Code is MIT for Python and Apache-2.0 for Rust, and the model weights are CC-BY-4.0, so commercial use is permitted with attribution, an unusually permissive stance for streaming speech models of this quality.