Turn handling

config.turn_handling controls the edges of a turn: whether the caller can cut the agent off, what happens after a false interruption, and whether the agent drafts a reply before the caller has certainly finished. Voice and video only; a text agent resets the block. Which end-of-turn detector runs is not configured here — it follows from the speech-to-text model and config.language. See speech and turn detection before tuning endpointing.

{
  "turn_handling": {
    "interruption": {
      "enabled": true,
      "min_speech_duration": 0.5,
      "min_words": 0,
      "resume_false_interruption": true,
      "false_interruption_timeout": 2.0,
      "discard_audio_if_uninterruptible": true
    },
    "preemptive_generation": {
      "enabled": true,
      "preemptive_tts": false,
      "max_speech_duration": 10.0,
      "max_retries": 3
    },
    "endpointing": { "min_silence_duration": 0.5, "max_silence_duration": 2.5 }
  }
}

Interruption

min_words needs interim transcripts. On a batch speech-to-text model it is a warning: the agent notices the interruption only once the caller stops speaking.
false_interruption_timeout: null disables false-interruption detection, so nothing is ever resumed.
A video agent always runs with resume_false_interruption off — the avatar cannot pause mid-sentence.
discard_audio_if_uninterruptible: false keeps what the caller said while the agent could not be interrupted and delivers it afterwards.

greeting_interruptible (top level, default true) controls the greeting. false makes the agent finish its opening line; it is refused on a realtime model. A greeting that speaks the recording notice is never interruptible.

Preemptive generation

The language model (and with preemptive_tts, the voice) starts on a reply before the turn is confirmed. It shortens the pause; a wrong guess is discarded and still billed on your provider key. On by default. It saves the most on a batch speech-to-text model, whose turns the VAD ends; a streaming model ends turns itself, leaving it little to overlap. It is forced off on a realtime agent (the model detects turns itself) and on an agent with vision_input.screenshare enabled (the context changes every turn, so every draft would be discarded).

Endpointing

min_silence_duration is how long the caller must stay silent before the agent takes the turn. max_silence_duration is the longer wait used when the turn detector judges the caller is mid-thought — only a batch speech-to-text model in one of the turn detector’s languages gets that judgement; elsewhere it is ignored. Must be ≥ min_silence_duration.

Symptom to setting

What you hear	Try
The agent cuts people off, or waits too long	Start at speech and turn detection, then `endpointing`
The agent stops for a cough or “mm-hm”	Raise `interruption.min_speech_duration`, or `min_words` 2–3 on a streaming model
The agent stops mid-sentence and never resumes	`resume_false_interruption` off or `false_interruption_timeout: null` (always off on video)
The agent resumes after a real interruption	Lower `false_interruption_timeout`
Long pause before replies	Enable `preemptive_tts`; check `reasoning_effort`
Preemptive generation has no effect	The agent is realtime or watches a screen share

Reference

Field	Type	Default	Range / notes
`interruption.enabled`	boolean	`true`	Off: the agent finishes every line
`interruption.min_speech_duration`	number	`0.5`	≥ 0. Seconds of speech that count as an interruption
`interruption.min_words`	integer	`0`	≥ 0. 0 = word count not consulted
`interruption.resume_false_interruption`	boolean	`true`	Forced off on video
`interruption.false_interruption_timeout`	number \| null	`2.0`	≥ 0. Seconds of silence before an interruption is judged false
`interruption.discard_audio_if_uninterruptible`	boolean	`true`
`preemptive_generation.enabled`	boolean	`true`	Forced off on realtime and screen share
`preemptive_generation.preemptive_tts`	boolean	`false`	Forced off on realtime and screen share
`preemptive_generation.max_speech_duration`	number	`10.0`	≥ 0. Longest caller speech it is attempted for
`preemptive_generation.max_retries`	integer	`3`	≥ 0. Attempts per user turn
`endpointing.min_silence_duration`	number	`0.5`	≥ 0.25
`endpointing.max_silence_duration`	number	`2.5`	≥ `min_silence_duration`
`greeting_interruptible` (top level)	boolean	`true`	`false` refused on realtime

Speech and turn detection

What ends a caller’s turn.

Debugging agents

When a call sounds wrong.

Get started

Build

Guides

Interruption

Preemptive generation

Endpointing

Symptom to setting

Reference

Next

Speech and turn detection

Debugging agents

​Interruption

​Preemptive generation

​Endpointing

​Symptom to setting

​Reference

​Next

Speech and turn detection

Debugging agents

Interruption

Preemptive generation

Endpointing

Symptom to setting

Reference

Next