A complete voice chat app powered by a speech-to-speech language model and bidirectional streaming.

On the backend is Kyutai Lab's Moshi model, which will continuously listen, plan, and respond to a the user. It uses the Mimi streaming encoder/decoder model to maintain an unbroken stream of audio in and out, and a speech-text foundation model to determine when and how to respond.

Thanks to bidirectional websocket streaming and use of the Opus audio codec for compressing audio across the network, response times on good internet can be nearly instantaneous, closely matching the cadence of human speech.

You can find the demo live here.

This repo is meant to serve as a starting point for your own language model-based apps, as well as a playground for experimentation. Contributions are welcome and encouraged!

[Note: this code is provided for illustration only; please remember to check the license before using any model for commercial purposes.]

1. React frontend (src/frontend/), served by src/app.py
2. Moshi websocket server (src/moshi.py)

• modal installed in your current Python virtual environment (pip install modal)
• A Modal account (modal setup)
• A Modal token set up in your environment (modal token new)

The Moshi server is a Modal class module to load the models and maintain streaming state, with a FastAPI http server to expose a websocket interface over the internet.

To run a development server for the Moshi module, run this command from the root of the repo.

modal serve src.moshi

In the terminal output, you'll find a URL for creating a websocket connection.

While the modal serve process is running, changes to any of the project files will be automatically applied. Ctrl C will stop the app.

From a seperate terminal, we can test the websocket connection directly from the command line with the tests/moshi_client.py client.

It requires non-standard dependencies, which can be installed with:

python -m venv venv
source venv/bin/activate
pip install -r requirements/requirements-dev.txt

With dependencies installed, run the terminal client with:

python tests/moshi_client.py

And begin speaking! Be sure to have your microphone and speakers enabled.

The http server at src/app.py is a second FastAPI app, for serving the frontend as static files.

A development server can be run with:

modal serve src.app

Since src/app.py imports the src/moshi.py module, this also starts the Moshi websocket server.

In the terminal output, you'll find a URL that you can visit to use your app. While the modal serve process is running, changes to any of the project files will be automatically applied. Ctrl C will stop the app.

Note that for frontend changes, the browser cache may need to be cleared.

Once you're happy with your changes, deploy your app:

modal deploy src.app

This will deploy both the frontend server and the Moshi websocket server.

Note that leaving the app deployed on Modal doesn't cost you anything! Modal apps are serverless and scale to 0 when not in use.