Barge-in teardown race

acknowledged

11 independent reportsacross 3 sourceson 9 stackslast reported 2026-08-20

What this is

When the caller interrupts the agent mid-speech, the system's interrupt handling fails to cleanly tear down the in-progress turn, corrupting the conversation state.

Status acknowledged: the thread it was reported in was acknowledged upstream.

Evidence

superMDguy2026-08-20assemblyaideepgramelevenlabslivekitsileroGitHub issuesgithub.com/…/6929

...in #5844, I created a load testing script that has a lot of artificial interruptions. During a recent test, it surfaced another error: text Error in main_task Traceback (most recent...

First seen 2026-08-20 · our copy fetched 2026-08-20T17:03:54Z · SHA-256 of that copy begins 5cd240b7f75bafe0… · profile

MahmoudHousam2026-08-15deepgrampipecatGitHub issuesgithub.com/…/5337

...same LLMSwitcher Feedback node sets LLMConfigureOutputFrame(skip_tts=True) so markdown is not spoken We also queue InterruptionWorkerFrame + flush_pipeline(timeout=2.0) before set_node_from_config Observed Bot spoke the last examiner question after the feedback node...

First seen 2026-08-15 · our copy fetched 2026-08-15T14:03:57Z · SHA-256 of that copy begins aa5b60aa445b367d… · profile

dtran262026-07-30livekitGitHub issuesgithub.com/…/2167

...callerHangupInstruction. Because that generation still has the consultation persona, briefing context, and the agent's interrupted turn, the model can finish or repeat the briefing before complying with the hangup instruction....

First seen 2026-07-30 · our copy fetched 2026-07-30T14:26:01Z · SHA-256 of that copy begins 51b46fa19ac8b928… · profile

kyle_shechtman2026-07-25livekitcommunity.livekit.iocommunity.…silent/1728

...speech handle remained current for approximately 293 seconds and was repeatedly targeted by subsequent interruptions. During that period, eight later LLM and TTS operations completed successfully, but there were no...

First seen 2026-07-25 · our copy fetched 2026-07-25T19:13:16Z · SHA-256 of that copy begins 567777915ec2711a… · profile

7 further reports of this mode, same five columns

arsenieciprian2026-07-24livekitGitHub issuesgithub.com/…/2108

Summary When the user interrupts the agent (barge-in) in a cascaded AgentSession using the Google plugin (Vertex AI), the truncated assistant message is committed to the chat context after...

First seen 2026-07-24 · our copy fetched 2026-07-25T17:50:27Z · SHA-256 of that copy begins 152a8e13216dffc9… · profile

xhdrq2026-07-21azure-speechGitHub issuesgithub.com/…/3056

...SIGSEGV / SEGV_ACCERR - Crashing frame: platform_memmove ← MicrosoftCognitiveServicesSpeech (GetModuleObject+…, stripped) - Observed while interrupting / restarting synthesis (barge-in style): stopSpeaking() followed quickly by startSpeakingSsml(...) on the same SPXSpeechSynthesizer At...

First seen 2026-07-21 · our copy fetched 2026-08-04T20:03:52Z · SHA-256 of that copy begins 04ea39c23559c56a… · profile

Raysharr2026-07-19livekitsileroGitHub issuesgithub.com/…/2065

Describe the bug In @livekit/agents 1.5.3, when a pipeline reply is interrupted before its playout has started, AgentActivity.mainTask can hang forever on the interrupted speech handle's generation future. currentSpeech...

First seen 2026-07-19 · our copy fetched 2026-07-25T19:11:42Z · SHA-256 of that copy begins cc6b78bc59ee42a3… · profile

thecarlmartin2026-07-12livekitopenai-realtimeGitHub issuesgithub.com/…/6391

Bug Description On a livekit-plugins-xai realtime call, an interruption can desync the plugin's local RemoteChatContext from xAI's server-side item list. Once that happens, every subsequent conversation.item.added is dropped, because...

First seen 2026-07-12 · our copy fetched 2026-07-26T22:06:27Z · SHA-256 of that copy begins fb48d33249c3cc8f… · profile

timofey-TK2026-06-01pipecatGitHub issuesgithub.com/…/4611

...single aggregation buffer. That buffer is only cleared by push_aggregation() (on LLMFullResponseEndFrame) and by handle_interruptions(). The turn-start path does not clear it: python async def handle_llm_start(self, : LLMFullResponseStartFrame): await self.trigger_assistant_turn_started()...

First seen 2026-06-01 · our copy fetched 2026-07-25T17:50:31Z · SHA-256 of that copy begins 4891230123214ad1… · profile

OlliTapio2026-03-27pipecattwilioGitHub issuesgithub.com/…/246

...function_name="save_notes", arguments={}, ) # Simulate: FunctionCallsStartedFrame adds orphan entry aggregator.function_calls_in_progress[call_1.tool_call_id] = None # Simulate: InterruptionFrame arrives — no FunctionCallCancelFrame emitted # aggregator.function_calls_in_progress still has {"call_1": None} # Turn 2: LLM...

First seen 2026-03-27 · our copy fetched 2026-07-25T17:50:33Z · SHA-256 of that copy begins c4f08a35b3de8b4a… · profile

R_Student2026-01-14openai-realtimetwilioStack Overflowstackoverf…on-speech-i

OpenAI Realtime Proper way to truncate a live streaming conversation on speech interruption Twilio media streams

First seen 2026-01-14 · our copy fetched 2026-07-24T20:54:01Z · SHA-256 of that copy begins d2ac04d9774ae9aa… · CC BY-SA 4.0 · profile

Quotes run to 30 words around the reported failure and link to the post.Add your call to this entry Take a quote down within 24 hours.

What is measurable

Signal	`hotato autopsy ./call.wav` maps to this mode as detector kind `overlap_while_agent_talking`. Printed as BARGE-IN: the caller took the floor and the agent kept speaking, reported with the overlap in seconds. Needs the caller and agent on separate channels.
Establishes	The seconds of overlap where the caller had taken the floor and the agent kept speaking, with the timestamp in the call where it started. What else it establishes, and what your own trace settles The same recording produces the same overlap on every run, so the number is comparable across builds. What your trace settles Your pipeline trace at that timestamp names what held the floor: the interrupt handler, the VAD threshold, or the realtime API's own turn logic. Whether the caller meant to interrupt is a judgement about the conversation; the measurement covers who held the floor and for how long.
Tier	Tier 1 · dual-channel audio: caller and agent separated at capture, so overlap and timing are facts of the recording, exact to the sample. hotato is MIT and runs locally, getting started.

Status history

Date	State	What moved it
2026-07-26	openacknowledged	issue closed (completed) evidence dated 2026-07-02

Mode	Status	Shares
Endpoint misdetection	acknowledged	reported on azure-speech; reported on deepgram; reported on elevenlabs; reported on livekit; reported on pipecat; reported on silero; reported on twilio; reported on github; reported on livekit_forum
Agent replies to own output	acknowledged	reported on assemblyai; reported on deepgram; reported on elevenlabs; reported on livekit; reported on silero; reported on github
Audio gate cooldown blocks endpoint finalization	acknowledged	reported on deepgram; reported on livekit; reported on openai-realtime; reported on pipecat; reported on silero; reported on github

Discussion

Reader discussion of this mode: what fixed it and what did not. The reports above are attributed and linked; a comment is a comment.

Open this discussion on GitHub

The record

Ids, digests, every link this entry stands on, and how to take a quote down

Atlas id	`VFA-2026-0001`
Mode id	`pat_d4645a0f0620c222f17fdbc4`
Content digest	`sha256:d4645a0f0620c222f17fdbc428d5e52274ae402e732d2d76a0b12913ed68d5b9`
First reported	`2026-01-14`
Last reported	`2026-08-20`
First seen by the atlas	`2026-07-15T10:59:40.675Z`
Last re-checked	`2026-08-20T17:02:40Z`
Reported on	assemblyaiazure-speechdeepgramelevenlabslivekitopenai-realtimepipecatsilerotwilio
Sources	GitHub issues, community.livekit.io, Stack Overflow
Scope	11 independent public reports across 3 sources. This entry states where the mode has been reported; where else it can occur is what the next report settles.
References	reporthttps://github.com/livekit/agents/issues/6929 · superMDguy · GitHub issues · 2026-08-20 reporthttps://github.com/pipecat-ai/pipecat/issues/5337 · MahmoudHousam · GitHub issues · 2026-08-15 reporthttps://github.com/livekit/agents-js/issues/2167 · dtran26 · GitHub issues · 2026-07-30 reporthttps://community.livekit.io/t/node-1-5-3-pre-playout-interruption-wedges-current-speech-later-llm-tts-succeed-but-the-call-stays-silent/1728 · kyle_shechtman · community.livekit.io · 2026-07-25 reporthttps://github.com/livekit/agents-js/issues/2108 · arsenieciprian · GitHub issues · 2026-07-24 reporthttps://github.com/Azure-Samples/cognitive-services-speech-sdk/issues/3056 · xhdrq · GitHub issues · 2026-07-21 reporthttps://github.com/livekit/agents-js/issues/2065 · Raysharr · GitHub issues · 2026-07-19 reporthttps://github.com/livekit/agents/issues/6391 · thecarlmartin · GitHub issues · 2026-07-12 reporthttps://github.com/pipecat-ai/pipecat/issues/4611 · timofey-TK · GitHub issues · 2026-06-01 reporthttps://github.com/pipecat-ai/pipecat-flows/issues/246 · OlliTapio · GitHub issues · 2026-03-27 reporthttps://stackoverflow.com/questions/79867418/openai-realtime-proper-way-to-truncate-a-live-streaming-conversation-on-speech-i · R_Student · Stack Overflow · 2026-01-14
Machine surfaces	feed.json feed.atom feed.schema.json atlas.llms.txt
Removal	If you wrote something quoted on this page and want it gone, one email to [email protected] does it. The removal path takes the quote, the handle and the link off this page, out of the feeds and out of the llms.txt within 24 hours, and tombstones the id so the entry cannot quietly return at a new address.