Turn detection & interruption

Two questions decide the rhythm of a voice conversation: when has the user finished their turn? (endpointing) and what happens when they talk over the agent? (interruption). Getting these right is most of what makes an agent feel responsive rather than laggy or rude — commit too early and you cut the user off, too late and it feels slow; ignore barge-in and it talks over the caller. The runtime gives you explicit control over both, with safe defaults, via Config.Turn (TurnOptions).

Endpointing: when the user's turn ends

Endpointing decides the moment the user turn is committed and generation starts. Set the mode on Config.Turn.Mode:

• VAD — end-of-speech from the voice-activity detector, plus an endpointing delay. The delay can be dynamic (longer after a short utterance, shorter after a long one) when enabled.
• STT — final transcripts plus the delay; no VAD required.
• Manual — the runtime never auto-commits; you call GenerateReply yourself. Right for push-to-talk or text.
• Auto (default) — VAD if a VAD is configured, else STT, else manual.

The endpointing delay is the pause the agent waits out before deciding the user really stopped (vs. taking a breath). Shorter = snappier but more likely to interrupt; longer = more patient (MinEndpointingDelay defaults to 500ms).

Barge-in: interrupting the agent

While the agent is speaking, incoming user speech can barge in. If the active SpeechHandle allows interruptions and the speech clears the configured thresholds (InterruptionMinDuration, InterruptionMinWords), the reply is cancelled, the audio buffer is cleared, and the session returns to listening.

A handle can opt out — a critical disclaimer can set itself uninterruptible so it always completes. A tool can do the same for its step with RunContext.DisallowInterruptions.

False interruptions

Not every sound is a turn. A cough, background speech, or a backchannel "mm-hm" can trip VAD without being a real interruption. The runtime treats these as a false interruption: rather than discard the reply, it pauses and resumes it after a short timeout if no real turn materialises — so a throat clear doesn't kill the agent's sentence.

Model-assisted end-of-turn (optional)

Beyond VAD/STT timing, you can plug in a model-assisted end-of-turn detector that predicts whether the user has really finished. It runs in an external Python sidecar over HTTP — the Go binary stays free of ML — and is entirely opt-in. inference.TurnDetector() builds one from the environment (INFERENCE_URL), returning nil (so endpointing falls back to plain VAD/STT) when unset:

session, _ := agents.NewSession(agents.Config{
    LLM:  "openai/gpt-4o",
    Turn: agents.TurnOptions{Detector: inference.TurnDetector()},
})

The moving parts

Next

• Speech & the turn — the generation half.
• Full API: Reference → TurnOptions.