xn-moshi

Rust implementation of the Moshi model family — the Mimi neural audio codec, the streaming ASR model, and the speech-to-speech model — built on top of the xn inference framework.

This crate used to live in the xn repository as the xn-moshi workspace member.

Usage

Add the dependency:

[dependencies]
xn-moshi = "0.2.4"

The library is backend-agnostic; pick one through the feature flags, which are forwarded to xn:

Examples

The moshi example needs the audio feature. It downloads the weights from the Hugging Face hub on first use.

# Encode an audio file to Mimi codes and decode it back.
cargo run --release --features audio --example moshi -- audio-to-audio input.wav -o output.wav

# Transcribe an audio file.
cargo run --release --features audio --example moshi -- asr input.wav

# Run the speech-to-speech model.
cargo run --release --features audio --example moshi -- s2s input.wav --voice voice.safetensors --config config.json

Add e.g. --features cuda to run on a GPU.

The quantize example converts an ASR safetensors checkpoint to a quantized GGUF file:

cargo run --release --example quantize -- asr model.safetensors model-q8.gguf --quant q8_0

Development

Install the formatting pre-commit hook with:

./scripts/setup-git-hooks.sh

License

Licensed under either of Apache License, Version 2.0 or MIT license at your option.