Voice agents

Turns, barge-in, and pushing text as your LLM writes it.

A turn is one thing your agent says — usually one LLM reply, often several sentences. You choose its context_id.

1. Open the socket and start

   One connection carries the whole conversation. Send start first, then read metadata.

   await ws.send(json.dumps({"type": "start", "v2": True,
                             "voice": "Ananya", "language": "hi"}))
   await ws.recv()
   
2. Push sentences as they are written

   Same context_id, continue: true until the last one. Do not wait between sentences — fire them as fast as your LLM produces them.

   await ws.send(json.dumps({"type": "text", "context_id": "t1",
                             "text": "आपका ऑर्डर कल पहुँच जाएगा।",
                             "continue": True}))
   await ws.send(json.dumps({"type": "text", "context_id": "t1",
                             "text": "कुछ और चाहिए?",
                             "continue": False}))
   
3. Cancel the moment the user speaks

   Barge-in drops the whole turn — queued and in-flight — and replies cancelled instead of end.

   await ws.send(json.dumps({"type": "cancel", "context_id": "t1"}))
   

The rules that matter

One end per turn, always

Closing a turn gives you exactly one end, tagged with its context_id. A cancelled turn gives you cancelled instead — never both.

context_id cannot be reused

Once a turn has ended, that id is spent. Reusing it is an error rather than a second turn, because two terminators for one id desynchronises everything after it.

Cancel before you close the socket

Closing alone leaves the GPU finishing audio nobody will hear. Send cancel, then close.

Audio arrives in order

Sentences are spoken in the order you sent them, tagged with the context_id, and streamed as they are generated.

Latency

Measured against Mumbai, eight turns on one socket:

Per-turn first audio	83 ms (79–95)
Handshake	269 ms, paid once