Interruption Settings

Control how your AI assistant handles caller interruptions — enable or disable barge-in, protect greetings, and tune endpointing thresholds for non turn-taking transcription models.

Interruption settings control whether and how callers can interrupt your AI assistant while it is speaking. By default, interruptions are enabled — callers can barge in at any point. You can disable interruptions entirely, protect only the greeting from being cut off, or tune the endpointing thresholds that determine when the assistant treats caller speech as a complete turn.

In this guide, you will learn:

• How interruption handling differs between turn-taking and non turn-taking transcription models.
• How to configure interruption settings via the API.
• How to protect greetings and tune endpointing thresholds.

---

How it works

Interruption behavior depends on which transcription model your assistant uses.

Turn-taking models

Turn-taking models like deepgram/flux have built-in end-of-turn detection. When you use one of these models, the assistant determines when the caller has finished speaking using the transcription-level settings eot_threshold, eot_timeout_ms, and eager_eot_threshold — configured under transcription.settings. See the Transcription Settings guide for details.

You can still use interruption_settings.enable and interruption_settings.disable_greeting_interruption with turn-taking models, but the start_speaking_plan and transcription_endpointing_plan thresholds are not relevant because the model handles turn detection natively.

Non turn-taking models

For models without built-in turn detection (such as deepgram/nova-3, deepgram/nova-2, azure/fast, or nvidia/parakeet-v3), the assistant relies on the interruption_settings.start_speaking_plan to decide when to start speaking after the caller stops. This plan uses silence-based endpointing thresholds to detect end of turn. Setting any start_speaking_plan — even an empty {}, one that matches the documented defaults, or one that carries only wait_seconds — also engages per-turn transcript endpointing, which can add or reduce per-turn wait relative to wait_seconds alone (see Presence of the plan changes behavior, not just its values).

---

Configuration

API schema

Set the interruption_settings object when creating or updating an assistant:

curl -X POST https://api.telnyx.com/v2/ai/assistants \
  -H "Authorization: Bearer $TELNYX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "My Assistant",
    "model": "anthropic/claude-haiku-4-5",
    "instructions": "You are a helpful voice assistant.",
    "transcription": {
      "model": "deepgram/nova-3"
    },
    "interruption_settings": {
      "enable": true,
      "disable_greeting_interruption": true,
      "start_speaking_plan": {
        "wait_seconds": 0.4,
        "transcription_endpointing_plan": {
          "on_punctuation_seconds": 0.1,
          "on_no_punctuation_seconds": 1.5,
          "on_number_seconds": 0.5
        }
      }
    }
  }'

Key fields

Field	Type	Default	Description
enable	boolean	true	Whether callers can interrupt the assistant while it is speaking.
disable_greeting_interruption	boolean	—	When true, prevents callers from interrupting the assistant’s greeting.
start_speaking_plan.wait_seconds	float	0.4	Wait used when per-turn endpointing does not select a tier: applied on every turn while the plan is present, and on turns where no tier matches (including when transcription_endpointing_plan is unset). Tier values replace it per turn — a tier may be shorter or longer.
start_speaking_plan.transcription_endpointing_plan.on_punctuation_seconds	float	0.1	Per-turn wait applied when the transcript ends with punctuation (., ?, !). Only evaluated while start_speaking_plan is present.
start_speaking_plan.transcription_endpointing_plan.on_no_punctuation_seconds	float	1.5	Per-turn wait applied when the transcript ends without punctuation or a number. Only evaluated while start_speaking_plan is present.
start_speaking_plan.transcription_endpointing_plan.on_number_seconds	float	0.5	Per-turn wait applied when the transcript ends with a number or a number word. Only evaluated while start_speaking_plan is present.

Presence of the plan changes behavior, not just its values

The plan’s thresholds only apply while interruption_settings.start_speaking_plan is set — but setting it engages a per-turn transcript matcher that does more than replace wait_seconds:

• Any present plan object engages per-turn matching — even an empty {} (the stored config fills in the field defaults) — including one that carries only wait_seconds, or one identical to the documented defaults. The thresholds above are then evaluated on each turn against the caller’s transcript tail.
• When the transcript matches a tier, the per-turn wait is that tier’s seconds outright: on_number_seconds for a transcript ending in a number or number word, on_punctuation_seconds when it ends in terminal punctuation (., ?, !), and on_no_punctuation_seconds otherwise. The tier value replaces wait_seconds for that turn — it is not clamped to it, so a tier can be shorter than wait_seconds.
• The no-punctuation tier always matches when transcription_endpointing_plan is set: a transcript with no terminal punctuation and no number falls through to on_no_punctuation_seconds (1.5 s by default). Per-turn waits therefore differ by configuration: with the tier plan set the wait is the tier value (1.5 s by default), with a present plan and no tier plan it is wait_seconds (0.4 s), and with no plan at all the assistant begins speaking as soon as silence is detected — present-tier configurations wait 1.5 s on every pause, which is that slower-than-absent behavior callers see.
• To turn per-turn matching off entirely, set start_speaking_plan to null (or omit it). The assistant then begins speaking as soon as the caller’s silence is detected, with no per-turn transcript evaluation. The nested tier plan is cleared separately: sending start_speaking_plan.transcription_endpointing_plan: null removes the tiers and the wait falls back to wait_seconds alone, while the outer plan stays engaged.

Sending a plan identical to the documented defaults still engages the matcher, so its per-turn values (0.1 / 1.5 / 0.5 by default) — not a flat wait — govern response latency on each turn. deepgram/flux assistants are the exception: ACA provisions their default plan with all three tiers at 0.1, so a Flux assistant’s plan behaves consistently whether sent explicitly or provisioned by default.

Mission Control Portal

You can also configure interruption settings through the Portal:

1. Navigate to AI > Assistants.
2. Select your assistant and click Edit.
3. Open the Voice tab.
4. Scroll to the Interruption Settings section.
5. Toggle interruptions on or off, enable greeting protection, and adjust the start speaking plan thresholds.

---

Example configurations

Default — allow interruptions

Allow callers to interrupt at any time, including during the greeting. This is the default behavior when interruption_settings is omitted.

{
  "interruption_settings": {
    "enable": true
  }
}

Protected greeting

Allow interruptions during normal conversation but prevent callers from cutting off the assistant’s initial greeting. Useful for flows where the greeting contains required disclosures or context.

{
  "interruption_settings": {
    "enable": true,
    "disable_greeting_interruption": true
  }
}

Conservative endpointing for non turn-taking models

Use longer silence thresholds to reduce false interruptions in noisy environments or when callers tend to pause mid-sentence. This configuration waits longer before treating silence as end of turn.

{
  "interruption_settings": {
    "enable": true,
    "start_speaking_plan": {
      "wait_seconds": 0.8,
      "transcription_endpointing_plan": {
        "on_punctuation_seconds": 0.3,
        "on_no_punctuation_seconds": 2.0,
        "on_number_seconds": 1.0
      }
    }
  }
}

Patient mid-sentence, fast after a finished sentence

Endpointing tiers key off the transcript tail, so a longer on_no_punctuation_seconds with a short on_punctuation_seconds waits for callers who pause mid-sentence, but responds quickly once the transcript ends in terminal punctuation:

{
  "interruption_settings": {
    "enable": true,
    "start_speaking_plan": {
      "wait_seconds": 1.5,
      "transcription_endpointing_plan": {
        "on_punctuation_seconds": 0.1,
        "on_no_punctuation_seconds": 2.0
      }
    }
  }
}

The tier values replace wait_seconds per turn rather than flooring it, so the example works as written: a transcript ending in ., ?, or ! waits on_punctuation_seconds (0.1 s), one that does not waits on_no_punctuation_seconds (2.0 s). The plan must stay present for the matcher to run, and wait_seconds (1.5 s here) governs only turns where transcription_endpointing_plan is unset or null (with the tier plan set, the no-punctuation tier always matches, so every turn gets a tier value).

---

Use cases

• Call centers with background noise: Increase wait_seconds and endpointing thresholds to prevent ambient noise from triggering false interruptions.
• IVR-style interactions: Keep interruptions enabled so callers can barge in to skip prompts they have already heard.
• Greeting-heavy flows: Enable disable_greeting_interruption to ensure callers hear required disclosures, legal notices, or welcome messages before interacting.
• Dictation or number entry: Increase on_number_seconds to give callers time to finish reading out long numbers like account IDs or phone numbers without the assistant cutting in.

---

Related resources

• Transcription Settings — Configure STT models and end-of-turn detection for turn-taking models like deepgram/flux.
• Create an Assistant API Reference — Full assistant configuration options including interruption_settings.
• End-of-Turn Detection — Lower-level end-of-turn parameters for WebSocket STT streaming.