fix(voice): enforce interruption min_duration on STT transcript triggers (#3515)

[dorukdumlu]

Closes #3515

Problem

interruption.min_duration (formerly min_interruption_duration) is only enforced on the VAD path: on_vad_inference_done checks ev.speech_duration >= min_duration before calling _interrupt_by_audio_activity(). The STT paths bypass it entirely — on_interim_transcript and on_final_transcript interrupt as soon as any non-empty transcript arrives, and _interrupt_by_audio_activity() itself only gates on min_words.

With providers that stream continuous interim results (Amazon Transcribe, etc.), the first interim lands well before min_duration is reached, so the option effectively does nothing — which is what was reported in #3515 (details in my comment there). This implements Option A from that comment: all entry paths share the same min_duration semantics.

Fix

• AudioRecognition gains a current_speech_duration property derived from the speech timestamps it already tracks (_speech_start_time / _last_speaking_time): wall-clock elapsed while the user is speaking, actual segment length once they've stopped, None when no speech start was ever tracked.
• _interrupt_by_audio_activity() enforces min_duration against that value, next to the existing min_words gate, so the interim/final transcript paths can no longer bypass it.

Two deliberate exemptions:

• Unknown duration (None) is allowed through. This preserves the existing failsafe in on_final_transcript ("agent speech might not be interrupted if VAD failed and a final transcript is received"), and keeps VAD-less sessions behaving as before rather than becoming un-interruptible.
• The interruption detection model path (on_interruption) skips the check via enforce_min_duration=False — that model already made an explicit "this is a real interruption" decision.

The VAD path keeps its existing pre-check; the new gate can't disagree with it (wall-clock elapsed ≥ VAD-measured speech duration), so VAD-triggered interruptions are unaffected.

Testing

• New regression test test_min_interruption_duration_applies_to_stt_transcripts: user speaks for 4s with interim transcripts streaming every 0.3s and min_duration=2.0. Before the fix the first interim interrupts at ~5.3s (playback position ~1.8s); after the fix the interruption fires at 7.0s (position 3.5s). Verified it fails without the fix and passes with it.
• FakeSTT/FakeActions gained an opt-in interim_interval to simulate continuous-interim providers; default behavior of existing fixtures is unchanged.
• Full --unit suite passes; ruff clean; no new mypy errors.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[dorukdumlu]

Friendly ping — CI is green on this one too. The change is intentionally small: gate STT-triggered interruptions on the same min_duration the VAD path already uses, with an escape hatch for the adaptive interruption model. Happy to tweak if you'd prefer a different call site.

[dorukdumlu]

Addressed Devin's finding on current_speech_duration:

Prefer VAD-measured speech_duration (same metric on_vad_inference_done already uses for min_duration) instead of wall-clock elapsed while _speaking. Wall-clock remains the fallback for STT-only setups with no VAD events.

Keeps short coughs/noise from clearing min_duration just because the interim transcript arrived after trailing silence.

[dorukdumlu]

Addressed the post-EOS Silero overwrite: after END_OF_SPEECH, Silero resets pub_speech_duration to 0, so later INFERENCE_DONE events were clobbering _vad_speech_duration and making late STT finals look too short for interruption.min_duration.

Now we only update from INFERENCE_DONE while still speaking or when speech_duration > 0, and keep the EOS final otherwise. Added a unit regression in tests/test_speech_start_time_persistence.py.

[dorukdumlu]

Addressed Devin's stale-duration finding: _bounce_eou_task now clears _vad_speech_duration with the other speech anchors on turn commit, so the next STT-failsafe path sees unknown duration (None) instead of the previous segment. Regression in ests/test_speech_start_time_persistence.py.

[dorukdumlu]

Addressed Devin's remaining yellow: _vad_speech_duration is now written before on_vad_inference_done, so _interrupt_by_audio_activity sees the same frame the VAD pre-check used. Post-EOS zero guard is unchanged. Regression in ests/test_speech_start_time_persistence.py.

[devin-ai-integration]

Devin Review found 1 new potential issue.

View 2 additional findings in Devin Review.

[devin-ai-integration]

🟡 Agent can't be interrupted when the microphone-based speech detector misses the user's speech

A zero speech length is recorded (self._vad_speech_duration = ev.speech_duration at livekit-agents/livekit/agents/voice/audio_recognition.py:1435-1436) while the user is already known to be talking, so the safety net that lets a finished transcript interrupt the agent is turned off and the agent keeps talking over the user.

Impact: In sessions where speech-to-text drives turn taking and the local speech detector fails to register the user's voice, barge-in stops working for that turn.

Mechanism: `_speaking` set by STT while VAD reports speech_duration == 0 turns "unknown duration" into "0.0 → too short"

The new min_duration gate in livekit-agents/livekit/agents/voice/agent_activity.py:2029-2036 deliberately lets an unknown duration (None) through so the VAD-failure failsafe in on_final_transcript (livekit-agents/livekit/agents/voice/agent_activity.py:2255) still works.

However, in turn_detection="stt" mode _speaking is set to True by the STT START_OF_SPEECH branch (livekit-agents/livekit/agents/voice/audio_recognition.py:1356-1365), independently of VAD. Silero only accumulates pub_speech_duration once pub_speaking is true, so every INFERENCE_DONE frame emitted while VAD has not (or never) detected onset carries speech_duration == 0.0. Because the guard is self._speaking or ev.speech_duration > 0.0, those frames write 0.0 into _vad_speech_duration, and current_speech_duration then returns 0.0 instead of None — the gate blocks the interruption permanently for that segment (0.0 < min_duration).

The self._speaking or disjunct is not needed for the stated purposes: post-EOS Silero zeros are already excluded by ev.speech_duration > 0.0, and any real voiced frame has speech_duration > 0.

Suggested change

	if self._speaking or ev.speech_duration > 0.0:
	self._vad_speech_duration = ev.speech_duration
	if ev.speech_duration > 0.0:
	self._vad_speech_duration = ev.speech_duration

---

Was this helpful? React with 👍 or 👎 to provide feedback.