Mimi

Mimi is Kyutai's state-of-the-art neural audio codec, developed as the audio backbone of the Moshi full-duplex speech dialogue system. It compresses speech into discrete tokens at 12.5 Hz and roughly 1.1 kbps while preserving prosody and speaker identity, combining semantic and acoustic information in a streaming encoder-decoder with a quantized latent space.\n\nIts low bitrate and causal streaming design make it the foundation for real-time speech language models: Moshi, Hibiki speech translation and Kyutai's STT/TTS models all tokenize audio through Mimi. It runs at 24 kHz mono and scales from 8 to 32 codebooks.\n\nReleased under CC-BY, Mimi is widely used in on-device and research voice AI, with community Rust, ONNX and mobile ports extending it beyond the official PyTorch stack.