Turn detection

What it does

English streams run turn-1-mini, our turn-detection model, on the incoming audio. At each pause it judges whether the speaker has finished or is only thinking. When it decides the turn is over, Audrey finalises the utterance at once, without waiting for endpoint_ms of silence, and then sends a turn_end message.

{"type": "final", "text": "Can you move it to Thursday?", "start": 4.1, "end": 5.86, "decode_ms": 132}
{"type": "turn_end", "probability": 0.912, "at": 6.12}

turn_end fields

How it behaves

• turn_end follows the final of the utterance it ends.
• If a pause of endpoint_ms finalised the utterance first, a turn_end can still follow during the same pause, without a new final.
• No turn_end follows an utterance you finalised with {"type": "flush"}.
• turn_end is informational: a voice agent can use it as the signal to reply. Transcripts need only the finals.

Turning it off

Turn detection is on by default for English. Send "turn_detection": false in the configuration to end utterances only on pauses and your flushes.

{"language": "en", "sample_rate": 16000, "turn_detection": false}