Turn detection (end-of-turn)

What it is

In conversation, people take turns. A voice agent has to guess when your turn ends, using pauses, words and tone, and when you’re interrupting it (“barge-in”).

How it works

• Silence-based (VAD). The agent waits for a pause of a set length. Simple, but pauses mid-sentence trigger early replies. OpenAI’s Realtime API defaults to this mode (server_vad).
• Meaning-based. The model judges from your words whether you’ve finished. OpenAI’s semantic_vad mode does this.
• Model-integrated. The speech model itself predicts the end of a turn. Deepgram’s Flux emits StartOfTurn, EagerEndOfTurn, TurnResumed and EndOfTurn events, with thresholds you can tune.
• Eager responses. Start preparing an answer on an early end-of-turn signal, and cancel it if the speaker resumes. In Flux, a TurnResumed event can only follow an EagerEndOfTurn.
• Barge-in. When the user starts talking, stop the agent’s audio immediately.

Benefits

• Faster, more natural replies.
• Fewer interruptions and talk-overs.
• Lower cost, because you don’t generate answers you’ll throw away.

Limits and risks

• Thresholds trade speed against false triggers, so tune them for your users.
• Echo from the agent’s own voice can trigger false turns. Use echo cancellation, or mute the input while the agent speaks.
• Options and names differ by vendor (checked 2026-10-17). IonBin doesn’t rank them. Test with real callers.