Barge-in

When a user speaks over your agent, you need to know quickly whether to stop. "Wait, no" is a claim on the floor; "mm-hm" is a backchannel and your agent should keep going.

Barge-in works on every /v2 endpoint. In a room it reacts to anyone in the room speaking over your agent.

1. Subscribe

   Add bargein to signals in session.start. If session.started.models.bargein is null, barge-in is unavailable for this session and no barge-in events will arrive.

   {
     "type": "session.start",
     "api_key": "baton_...",
     "eot": {"profile": "cutoff_5pct"},
     "signals": ["vad", "eot", "bargein"]
   }
   

2. Tell Baton when your agent talks

   {"type": "agent.speech", "phase": "start",
    "utterance_id": "sp_1"}
   {"type": "agent.speech", "phase": "end",
    "utterance_id": "sp_1"}
   

   Messages are timestamped on arrival against the audio received so far. phase: "end" drops an interruption that has not been classified yet.

3. Act on the classification

   {"type": "bargein.onset", "seq": 7, "t": 12.1,
    "agent_speaking": true, "utterance_id": "sp_1"}
   {"type": "bargein.class", "seq": 8, "t": 12.4,
    "class": "claim", "p_claim": 0.9, "decided_after_ms": 300,
    "utterance_id": "sp_1"}
   

   On claim, stop speaking and listen. On backchannel, keep going.

How it decides

• The onset is the start of user speech while your agent is speaking.
• Baton classifies the interruption as soon as it is confident, typically a few hundred milliseconds in. decided_after_ms reports how long it took.
• p_claim is the confidence that the user is claiming the floor. It is null when the user stopped too soon to judge; that case is reported as a backchannel.