This repo contains instructions and examples of how to run Kyutai Speech-To-Text models. These models are powered by delayed streams modeling (DSM), a flexible formulation for streaming, multimodal sequence-to-sequence learning.

Text-to-speech models based on DSM coming soon! Sign up here to be notified when we open-source text-to-speech and Unmute.

More details can be found on the project page.

Kyutai STT models are optimized for real-time usage, can be batched for efficiency, and return word level timestamps. We provide two models:

• kyutai/stt-1b-en_fr, an English and French model with ~1B parameters, a 0.5 second delay, and a semantic VAD.
• kyutai/stt-2.6b-en, an English-only model with ~2.6B parameters and a 2.5 second delay.

These speech-to-text models have several advantages:

• Streaming inference: the models can process audio in chunks, which allows for real-time transcription, and is great for interactive applications.
• Easy batching for maximum efficiency: a H100 can process 400 streams in real-time.
• They return word-level timestamps.
• The 1B model has a semantic Voice Activity Detection (VAD) component that can be used to detect when the user is speaking. This is especially useful for building voice agents.

You can retrieve the sample files used in the following snippets via:

wget https://github.com/metavoiceio/metavoice-src/raw/main/assets/bria.mp3
wget https://github.com/kyutai-labs/moshi/raw/refs/heads/main/data/sample_fr_hibiki_crepes.mp3

This requires the moshi package with version 0.2.6 or later, which can be installed via pip.

python -m moshi.run_inference --hf-repo kyutai/stt-2.6b-en bria.mp3

If you have uv installed, you can skip the installation step and run directly:

uvx --with moshi python -m moshi.run_inference --hf-repo kyutai/stt-2.6b-en bria.mp3

It will install the moshi package in a temporary environment and run the speech-to-text.

Additionally, we provide two scripts that highlight different usage scenarios. The first script illustrates how to extract word-level timestamps from the model's outputs:

uv run \
  scripts/streaming_stt_timestamps.py \
  --hf-repo kyutai/stt-2.6b-en \
  --file bria.mp3

The second script can be used to run a model on an existing Hugging Face dataset and calculate its performance metrics:

uv run scripts/streaming_stt.py  \
  --dataset meanwhile  \
  --hf-repo kyutai/stt-2.6b-en

The Rust implementation provides a server that can process multiple streaming queries in parallel. Dependening on the amount of memory on your GPU, you may have to adjust the batch size from the config file. For a L40S GPU, a batch size of 64 works well and requests can be processed at 3x real-time speed.

In order to run the server, install the moshi-server crate via the following command. The server code can be found in the kyutai-labs/moshi repository.

cargo install --features cuda moshi-server

Then the server can be started via the following command using the config file from this repository. For kyutai/stt-1b-en_fr, use configs/config-stt-en_fr.hf.toml, and for kyutai/stt-2.6b-en, use configs/config-stt-en-hf.toml,

moshi-server worker --config configs/config-stt-en_fr-hf.toml

Once the server has started you can run a streaming inference with the following script.

uv run scripts/asr-streaming-query.py bria.mp3

The script limits the decoding speed to simulates real-time processing of the audio. Faster processing can be triggered by setting the real-time factor, e.g. --rtf 500 will process the data as fast as possible.

A standalone Rust example script is provided in the stt-rs directory in this repo. This can be used as follows:

cd stt-rs
cargo run --features cuda -r -- bria.mp3

You can get the timestamps by adding the --timestamps flag, and see the output of the semantic VAD by adding the --vad flag.

MLX is Apple's ML framework that allows you to use hardware acceleration on Apple silicon.

This requires the moshi-mlx package with version 0.2.6 or later, which can be installed via pip.

python -m moshi_mlx.run_inference --hf-repo kyutai/stt-2.6b-en-mlx bria.mp3 --temp 0

If you have uv installed, you can skip the installation step and run directly:

uvx --with moshi-mlx python -m moshi_mlx.run_inference --hf-repo kyutai/stt-2.6b-en-mlx bria.mp3 --temp 0

It will install the moshi package in a temporary environment and run the speech-to-text.

The MLX models can also be used in swift using the moshi-swift codebase, the 1b model has been tested to work fine on an iPhone 16 Pro.

We're in the process of open-sourcing our TTS models. Check back for updates!

The present code is provided under the MIT license for the Python parts, and Apache license for the Rust backend. The web client code is provided under the MIT license. Note that parts of this code is based on AudioCraft, released under the MIT license.

The weights for the speech-to-text models are released under the CC-BY 4.0 license.