Skip to content

Live camera and microphone

How to turn on the WebSocket that carries camera frames and microphone audio, and what changes when it is on.

Live camera

A Camera component sends webcam frames to a handler you write. By default the frames go as one snapshot a second over the upload route, which works everywhere, Colab included. For a smooth feed, turn on the WebSocket:

pip install "indah[live]"

# live=indah.Live(max_sockets_per_address=2) tunes the caps
app = indah.create_app(session_factory=build, live=True)

Live is off unless you ask for it, and indah never turns it on for you. Colab's proxy blocks WebSockets, so there live=True is ignored with a warning and the camera sends snapshots. What changes when it is on:

  • The upgrade is refused for a cross-site Origin, whether or not the app is secure (a browser does not apply the same-origin policy to a WebSocket). allowed_origins applies.
  • On a secure app the socket uses the same signed session token as every other request.
  • Behind a proxy, set trusted_proxies as usual, and make sure the proxy passes Upgrade requests (Cloudflare, Fly and Traefik do).
  • With csp="strict", connect-src also names the page's own ws:// or wss:// origin. Nothing else about the policy changes, and with live off it is exactly as before.
  • Each address may hold 4 sockets, a frame over 1 MiB closes the socket, and a handler that is slower than the camera sees fewer frames, always the newest. Run a heavy model with await asyncio.to_thread(model, frame.data).
  • The browser only gives out the camera on https:// pages or localhost.

Live microphone

A Microphone rides the same socket and the same switches: live=True turns it on, and nothing is added for an app that does not. Its handler gets an AudioChunk of 16-bit mono PCM (about 100 ms), in order. If the handler is slower than speech, audio waits up to Live.audio_buffer_seconds (2 s), then the oldest is dropped and the next chunk says so in dropped_before. The page loads a small AudioWorklet from /api/live-audio.js on your own origin, which the strict CSP already allows. On Colab, or wherever the socket is unavailable, the browser records and sends one WAV take when you press Stop. See examples/mic_transcribe.py.