# Stop talking when interrupted, keep talking through backchannels

> Stop talking when interrupted, keep talking through backchannels: 1 companies fully do it, 16 with a limit, 30 carry the tag. Every listing in Conversational voice agents with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/voice-agents.interruption-handling/
- Markdown: https://www.anchorterminal.com/niches/voice-agents.interruption-handling/index.md (~2,250 tokens)
- Slim: https://www.anchorterminal.com/niches/voice-agents.interruption-handling/index.min.md (~480 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/voice-agents.interruption-handling/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

One company's docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 1 | 16 | 30 | 10 | 4 |

## What counts as done

Agent audio stops on a real interruption, backchannels under two words do not stop it, a marked utterance plays to the end, and the transcript records what the caller heard.

Measured, for our own tests: agent audio stops within 300 ms of a real interruption.

How agents ask for it: why does the agent keep talking over me; it stops every time I say mm-hmm; the consent line must play in full

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Vapi API + MCP (Vapi Inc.)**: stopSpeakingPlan has numWords, voiceSeconds and default acknowledgementPhrases (mm-hmm, right) that never interrupt; the controlUrl say command plays with interruptionsEnabled false by default; Vapi hosts the telephony leg.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Amazon Nova 2 Sonic (Amazon Web Services, Inc.)**: Barge-in stops generation and sends an interruption signal, and a post-response text gives what was spoken. The client must clear its own audio queue; no backchannel filter or uninterruptible utterance is documented.
- **AssemblyAI Voice Agent API (AssemblyAI, Inc.)**: Semantic barge-in by default: backchannels like uh-huh don't interrupt, and transcript.agent is trimmed to what the user heard. interrupt_response is session-level only; no way to mark one utterance uninterruptible.
- **Bland AI API + MCP (Intelliga Corporation)**: interruptibility 0 to 4 (default 2; only 4 stops on uh-huh) and block_interruptions are call-level. No documented way to mark one utterance uninterruptible while the rest stays interruptible.
- **Bolna API + MCP (Voxlabs Private Limited)**: An interruption word count filters backchannels, and number_of_words_for_interruption 0 makes a graph node uninterruptible. An open issue reports a TTS socket closing mid-reply leaving the agent marked as speaking.
- **Deepgram Voice Agent API (Deepgram, Inc.)**: Agent audio is held until the turn is confirmed and the Twilio guide clears Twilio's buffer on UserStartedSpeaking. No backchannel filter or per-utterance uninterruptible flag is documented.
- **ElevenLabs Agents API + MCP (Eleven Labs Inc.)**: Interruptions can be disabled at agent level and per tool (disable_during_tool_and_turn protects the following reply), with turn eagerness settings. No documented backchannel filter, and marking one arbitrary utterance uninterruptible is not shown.
- **Gemini Live API (Google LLC)**: Server VAD cancels generation and reports interrupted, with start and end sensitivity settings, but NO_INTERRUPTION is session-wide, not per utterance, and no backchannel filter is documented.
- **GPT-Live (OpenAI OpCo, LLC)**: Full duplex, so caller and assistant can overlap and the model handles turns, but no turn, backchannel or uninterruptible-utterance setting is documented; the guides tell the client to clear queued audio itself when speech is cancelled.
- **Inworld Realtime API (Theai, Inc. dba Inworld AI)**: interrupt_response cancels on speech, the Twilio example clears the carrier buffer, and response.speak takes force_uninterruptible for a fixed consent line. Caller backchannels are not filtered beyond the Silero threshold; the backchannel feature is the agent's own interjections.
- **LiveKit Agents (LiveKit Incorporated)**: Adaptive interruption mode filters backchannels and say(allow_interruptions=False) protects one line, but under a realtime model's server-side turn detection that flag is ignored and the request was closed not planned.
- **OpenAI Realtime API (OpenAI OpCo, LLC)**: Server VAD or semantic_vad cancels the response and WebRTC and SIP truncate unplayed audio on the server, but interrupt_response is a session setting, not per utterance, and no backchannel filter is documented.
- **Pipecat (Daily)**: MinWordsUserTurnStartStrategy and a Krisp model filter backchannels and mute strategies block barge-in, but mute strategies are session-wide or event-based rather than one marked utterance, and Twilio clear-on-interrupt is not stated in its serializer page.
- **Retell AI API + MCP (Retell AI Inc.)**: interruption_sensitivity (0 to 1) per agent or node and no_interruption_allowed per response on the custom-LLM socket protect a line, but no caller-backchannel filter is documented; Enable Backchanneling is the agent's own acknowledgements.
- **Rime Prism (Rime Labs, Inc.)**: Interruption is on by default and conversation.item.truncate records what the caller heard, but interrupt_response is session-wide, not per utterance, and no backchannel filter is documented in the pages read.
- **Ultravox Realtime API (Theai, Inc. dba Inworld AI)**: vadSettings.minimumInterruptionDuration filters short speech by duration and a clear-buffer message exists for WebSocket media, but only the first greeting can be marked uninterruptible.
- **Vaani Developer Platform (Vaani AI Research Private Limited (Bengaluru, India, CIN U62099KA2025PTC201212))**: Agents have an interruptible flag and welcome_interruptible per call, plus eagerness and patience settings, but no backchannel filter is documented and only the welcome line can be made uninterruptible per utterance.
- **Vocily AI (VOCILY AI PRIVATE LIMITED)**: interruption_sensitivity runs from 0 (never yields) to 1 (yields at any sound) and response_eagerness tunes turn-taking, but there is no backchannel list or per-utterance uninterruptible flag.
- **Vogent API (Monoid, Inc.)**: utteranceDetectorConfig has a sensitivity and interruptionPlayTimeMs (minimum time the agent keeps speaking after an interruption), and endpoint detection can be semantic, but there is no backchannel list or per-utterance uninterruptible flag.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **fone.me (Fone)**: No page that loads without JavaScript describes barge-in or backchannel handling.
- **Fresha AI Concierge (Fresha.com SV Ltd)**: No documentation beyond the product page, which says nothing about barge-in or backchannels.
- **Harmony.ai (monday.com Ltd.)**: No documentation of barge-in, backchannels or uninterruptible lines; only marketing copy about response speed.
- **Hume EVI (Empathic Voice Interface) (Hume AI, Inc.)**: EVI docs withdrawn: every EVI page, including the overview and tool-use guide, redirects to the Audio Decisions overview, and the EVI AsyncAPI file returns 404. The last assessment recorded EVI access ending 13 November 2026.
- **PlaceCall (VOYGR Tech, Inc.)**: The managed call agent exposes no interruption or turn settings and the docs say nothing about barge-in behaviour; no first-hand account.
- **Synthflow API + MCP (AgentFlow AI GmbH)**: The voice and call configuration pages describe patience level and noise cancellation but no interruption, backchannel or uninterruptible-message setting; no account covers it.

- not this job: CueLight (sshah03 (individual GitHub account, no name or organisation stated)): It predicts when the user has finished a turn; it does not stop agent audio, filter backchannels or protect an utterance. A pipeline would need its own barge-in logic.
- not this job: Pipecat PhoneLLM (Daily): A text model for the LLM step only; it has no audio, telephony or session layer, so the framework around it (Pipecat or another) decides this.

## What stops agents finishing

- server-side turn detection ignores allow_interruptions=False
- any detected speech counts as an interruption
- TTS socket closing mid-reply leaves the agent marked as speaking

## Around it

After: voice-agents.tool-call-mid-call.

## Where this comes from

Weight 10 of 100 in Conversational voice agents: 4 of 31 accounts (acc-02, acc-04, acc-05, acc-12) and one open barge-in scenario pack covering Vapi, Retell, Bland, Pipecat and LiveKit

As data: https://www.anchorterminal.com/api/v1/use-cases/voice-agents.interruption-handling.json, or the MCP tool get_use_case with id voice-agents.interruption-handling. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
