Voice pipeline engineering
How to reduce false triggers in voice AI
Stop firing on speech that was never meant for the device.
Root cause
False triggers are a routing error, not a transcription error.
Every utterance that clears VAD runs the full STT, LLM, TTS pipeline, meant for the device or not. The fix: a gate that answers the question VAD never asks.
VAD: is someone speaking right now. Selective Auditory Attention (SAA): is the device being spoken to right now.
Common approaches and their limits
| Approach | When it runs | Gates ASR input? | Decides addressee on natural speech? |
|---|---|---|---|
| Wake-word tuning | Pre-ASR | Partially | No, requires trigger phrase; natural speech without phrase is not gated |
| ASR confidence threshold | Post-ASR | No, ASR already ran; cost already incurred | Yes, but high confidence does not mean speech was device-directed |
| Speaker diarization | Post-ASR | No, runs after transcription | Identifies who spoke, not who was addressed |
| SAA addressee detection | Pre-ASR, in real time | Yes, only addressed audio reaches ASR | Yes, no trigger phrase or enrollment required |
Environments
False triggers occur wherever background speech reaches the microphone.
Multi-person environments see this most, but any ambient audio, a TV, a nearby conversation, produces the same failure at lower intensity.
Drive-thru
Crew cross-talk and multi-lane audio
Order-taking agents face crew cross-talk, adjacent lane bleed, and customers addressing companions instead of the speaker.
At drive-thru volume, that avoidable pipeline cost adds up fast.
Meeting rooms
Multi-party sessions with shifting speakers
A voice agent in a conference room hears every participant.
Without an addressee gate, it cannot distinguish a direct question from a side conversation or interjection.
Car cabins
Driver-to-passenger speech
In-vehicle assistants routinely activate on driver-to-passenger exchanges.
Addressee detection distinguishes a command to the vehicle from a remark to a companion.
How SAA works
Gate addressee before ASR. Forward only addressed audio. Fail closed when uncertain.
SAA sits between VAD and ASR, returning one addressed-or-not decision per utterance.
pre-ASR gate
Decision before transcription runs
SAA returns its addressee routing signal before ASR begins. Unaddressed speech never enters the transcription pipeline: no wasted token spend, no spurious response.
no trigger phrase
Natural speech, no enrollment required
SAA evaluates each utterance on audio features (and optionally video) without a trigger phrase or enrolled profile, covering the bystander cases wake words miss.
fail-closed design
Uncertainty suppresses, never passes through
SAA suppresses ambiguous cases instead of forwarding them. Uncertainty defaults to non-addressed, never to addressed.
latency
In real time, no GPU required
Fits live-session latency budgets. SAA runs as a hosted cloud endpoint; integration is a stream-intercept before your existing ASR call.
Accuracy
Measured on held-out multi-party sessions.
Full method and evaluation published at arXiv:2604.08412.
audio and video fusion
Highest accuracy on the addressee decision
Video resolves overlapping speech and off-axis talkers that audio alone leaves ambiguous.
audio-only
Reliable on audio alone
No camera needed. Runs as hosted cloud inference wherever video isn't available.
Every downstream filter treats a symptom. Pre-ASR addressee gating removes the cause.
Common questions
What voice AI builders ask about false triggers and addressee detection.
What is a false trigger and why does my voice agent respond to bystanders?
A false trigger is any activation caused by speech not directed at the device.
The root cause: voice activity detection tells the pipeline someone is speaking but not whether the device is being addressed.
A bystander, a crew member talking to a colleague, a passenger speaking to the driver: all clear VAD the same way a device-directed utterance does.
Can I fix false triggers by tuning ASR confidence thresholds or wake words?
Only partially. Confidence thresholds run post-ASR: the transcription already happened and the cost is incurred. A bystander speaking clearly can score high confidence.
Wake words need a trigger phrase, force unnatural interaction, and still misfire when bystanders say the phrase.
Neither approach decides addressee. Pre-ASR addressee detection removes the root cause.
Does SAA require a wake word or speaker enrollment?
No. SAA works on natural speech with no trigger phrase and no enrolled profiles.
It decides in real time, before STT, whether the current utterance is directed at the device, based on audio features (and optionally video).
It fails closed when uncertain: nothing is forwarded if the addressee decision is ambiguous.
How accurate is SAA at addressee detection?
SAA decides addressee reliably, in real time and before STT, on held-out multi-party sessions, using audio alone or audio plus video.
Full method and evaluation: arXiv:2604.08412.
It fails closed under distribution shift, suppressing uncertain cases rather than forwarding them. Cross-lingual recall is a known limitation under active work.
Get in touch
Request an eval or book a call.
Tell us about your voice stack and where you are deploying. We will follow up with eval next steps on your audio. Eval is free; licensing is per device.
Last updated June 29, 2026