Detect livestream activity on Chalk GPUs
This demo shows how to run a video and audio pipeline on a Chalk GPU to identify active moments and breaks in stream recordings. Play the sample to try it, then explore the code to see how it works.
Compare the recording with the detected breaks
The sample has known active and break segments. Check the detector against those moments.
Try this experiment · expected results
1Play the sample
Watch a talking segment, then jump to the streamer stepping away or the break screen.
Expected: Compare the detected strip with the script strip; the confidence line explains the changes.
2Inspect a disagreement
Click near a transition and compare person, scene, speech, and motion signals.
Expected: A person being visible is only one input; the combined verdict can lag a transition.
Chalk serves the video and audio models. Compare the detected breaks with the recording to see how several weak signals combine into one activity decision.
The streamer is fictional: generated with LTX-Video and voiced with Kokoro TTS. The sample analysis is computed at startup and plays in sync with the recording. Uploads are limited to 30 seconds and 15 MB; the app deletes its temporary upload file after analysis.
Built on Chalk
Chalk runs the models and web app on an NVIDIA L4. The example pipeline combines person detection, CLIP, voice activity, and motion into an activity score.
A model ready to use
Deploy the model and its web app together. Chalk provisions the GPU and returns an HTTPS URL.
Deploy from Python
Define the container and its model weights in Python. Later builds can reuse cached image layers.
Vision weights ready at startup
The image build downloads the detector and CLIP weights. New replicas load those weights from the image before analyzing the sample recording.
Signals you can inspect
The example’s fusion and smoothing logic turns four signals into active and break segments. The app shows those signals alongside each verdict so you can inspect the result.
Run it on your own
Clone the example, log in to Chalk, and run the deploy script. Chalk builds the image with detector and CLIP weights, provisions an NVIDIA L4, and prints the demo URL after the sample recording is analyzed.
You will need uv, the Chalk CLI, and a Chalk environment with GPU capacity. The upload API returns per-sample signals, activity probabilities, and segments. Adapt the pipeline and its fitted weights for your own recordings.
git clone https://github.com/chalk-ai/activity-detection
cd activity-detection
# Or export CHALK_CLIENT_ID and CHALK_CLIENT_SECRET.
chalk login
# Builds the image, provisions an L4, analyzes the sample stream, prints the URL.
./deploy.py --gpu nvidia-l4Find the active moments in your video
Bring video and audio models together for stream analysis, moderation, or highlight detection. Talk to an engineer about running your pipeline on Chalk GPUs.