Serve open speech-to-text models on Chalk GPUs
This demo shows how to run Kyutai’s open-source speech model on a Chalk GPU, turning your voice into a live transcript. Speak into your microphone to try it, then explore the code to see how it works.
Start talking
Press Start talking and allow microphone access. Speak in English or French.
Watch the transcript appear as you speak. Press Stop when you are done.
Start talking to stream audio to the GPU model served on Chalk. Compare the words you say with the transcript arriving while you speak.
This shared demo serves one speaker at a time. If it is busy, try again in a moment.
Built on Chalk
Chalk builds the model container, provisions the GPU, and serves the streaming endpoint. Connect your audio client to turn speech into live captions or power a voice interface.
A model ready to use
Deploy the model and its web app together. Chalk provisions the GPU and returns an HTTPS URL.
Streaming model endpoints
Chalk’s GPU endpoints support two-way streaming over WebSockets. Stream audio to the model and receive words as they’re generated, so your application can respond while someone is still speaking.
Deploy from Python
Define the container and its model weights in Python. Later builds can reuse cached image layers.
Model weights ready at startup
Chalk builds the model weights into the container image. New replicas load them from the image without downloading them again.
Run it on your own
Install uv and the Chalk CLI, then select an environment that can schedule an NVIDIA L4. Run Quickstart from the cloned repository; uv manages the scripts’ Python version and dependencies.
The first build can take 10–20 minutes. Open the printed web_url and select Start talking. The Test a recording tab uses the repository’s client and a public sample without a microphone. Each replica accepts one active stream at a time.
git clone https://github.com/chalk-ai/streaming-speech-to-text
cd streaming-speech-to-text
chalk login
chalk environment YOUR_ENVIRONMENT
# Builds the image and deploys one NVIDIA L4. First build: about 10–20 minutes.
uv run --no-project --python 3.12 --script deploy.py
# Open the printed web_url and select Start talking.Build a voice application on Chalk
Talk to an engineer about serving speech models and streaming audio in your Chalk environment.