mimi-rs

A Rust implementation of an audio codec that tokenizes and reconstructs speech.

Rust implementation of Kyutai’s Mimi audio codec, built on candle. Provides a streaming SEANet encoder/decoder, KV-cached transformer, RVQ quantizer, and post-load F32 to Q8_0 weight quantization. Shared library powering audio tokenization in tts-web and stt-web’s WASM pipelines.

Runs
browser, WASM, native
Model
kyutai/mimi, 96.2M params, CC-BY-4.0
Source
idle-intelligence/mimi-rs
Status
maintained

← Home