Docs
Realtime
Exchange audio over WebSocket for uninterrupted conversation, including barge-in while the user is still speaking.
Connecting
Connect over wss and send session configuration in the first frame. Auth goes in a query token rather than a header.
Events
Audio deltas, text deltas, and speech start/stop events stream in. Barge-in is handled with a cancel event.
| Event | Direction | Note |
|---|---|---|
| session.update | client → server | Voice, model, instructions |
| input_audio.append | client → server | PCM chunk |
| response.audio.delta | server → client | Output audio chunk |
| response.text.delta | server → client | Transcript delta |
| response.cancel | client → server | Barge-in |
Billing
Input and output audio are counted per minute. Holding a connection open without speaking is not billed.