End of turn · Baton docs

How Baton decides a speaker has finished, and how to trade speed against false cut-offs.

People pause mid-sentence all the time. Baton continuously judges whether the speaker is done, and sends turn.end only when it is confident they have finished.

How a turn ends

turn.end arrives after the speaker goes quiet, in one of two ways:

Baton is confident the turn is complete (by: "score"). p_eot is that confidence.
The silence has gone on long enough that the turn ends anyway (by: "timeout"). The limit depends on the operating point.

If the speaker starts again within 2 s of turn.end, Baton sends turn.resumed. Treat that turn end as cancelled.

Operating points

Each operating point is a published trade-off between speed and false cut-offs (pauses wrongly ended as turns), measured on Baton's end-of-turn benchmark.

Operating point	`/v2/stream`	`/v2/room`	False cut-offs	Good for
`cutoff_10pct`	350 ms	439 ms	≈ 10 %	Quick back-and-forth: bookings, support calls
`cutoff_5pct`	577 ms	726 ms	≈ 5 %	Conversations and meetings
`cutoff_2pct`	875 ms	1134 ms	≈ 2 %	People who think aloud: interviews, brainstorms

Latency is the mean delay from the end of speech to turn.end. English values; points for all 14 supported languages are published at GET /v2/eot/profiles.

Choosing one

Set eot in session.start, by name or by the latency you can afford. Without eot, a session uses cutoff_10pct.

By name

{
  "type": "session.start",
  "api_key": "baton_...",
  "eot": {"profile": "cutoff_5pct"}
}

By latency budget

{
  "type": "session.start",
  "api_key": "baton_...",
  "eot": {"latency_budget_ms": 600}
}

A latency budget picks the published point with the closest mean latency; ties go to the faster one. With {"latency_budget_ms": 600}, /v2/stream gets cutoff_5pct (577 ms) and /v2/room gets cutoff_5pct (726 ms); with 500, a room gets cutoff_10pct (439 ms).

session.started reports the point in use:

{
  "type": "session.started",
  "schema_version": 2,
  "session_id": "3f7c...",
  "tier": "full",
  "eot": {
    "profile": "cutoff_5pct",
    "mean_latency_ms": 577,
    "cutoff_rate": 0.0482
  }
}

Conversation context

On /v2/stream, pass the conversation so far in messages. Baton uses it to judge whether an utterance is complete: "Paris" after "Where are you flying?" is a full answer.

{
  "type": "session.start",
  "api_key": "baton_...",
  "messages": [
    {
      "role": "assistant",
      "content": "What city are you flying to?"
    }
  ]
}