Principal-speaker gating for drive-thru voice ordering
Your drive-thru AI takes orders from the wrong person.
Selective Auditory Attention (SAA) is the hosted layer that decides who the lane is addressing, before ASR runs, so only the driver reaches your stack.
The failure
The lane mic hears everyone. Your AI can't tell who's ordering.
The driver, the back seat, the car radio, the next lane, and the crew all reach the mic at once, with no signal for who is placing the order.
The recognizer isn't broken. Nothing decides whether an utterance was meant for the lane before transcription runs.
false trigger
The car answers for the customer
A back-seat joke, a loud radio, or the next lane ordering two feet away all read as speech to the lane AI, which adds items nobody asked for.
false rejection
The real correction gets buried
A fix like "no onions" can get lost in the same crosstalk, so a crew member has to override the ticket by hand while the line backs up.
The fix
One layer that decides who the lane is addressing.
SAA is a hosted, pre-ASR layer that decides in real time whether lane speech is addressed to the ordering device or to someone else in the car.
It routes only addressed audio to the ASR and model you already run, fails closed when unsure, and needs no wake word or GPU.
Research foundation
Built on our published technical paper.
SAA decides whether speech is addressed to the ordering device, using audio-video fusion or audio alone, across held-out multi-party sessions (arXiv:2604.08412).
Two caveats apply to both modes: it fails closed under distribution shift, and cross-lingual recall is a known limitation under active work.
performance
Highest accuracy with audio and video
Held-out multi-party sessions. Published at arXiv:2604.08412.
performance
Accurate on audio alone
Real multi-speaker audio, no video stream. The conservative floor for open-mic, audio-only environments.
latency
Real-time decision
Returned before your ASR starts processing, so it adds no latency to the lane.
design principle
Fails closed by design
Ambiguous speech is suppressed. The default under uncertainty is silence, not a false order.
coverage
Up to 4 simultaneous speakers
The multi-party load a full car plus lane crosstalk produces.
signal model
Acoustic and prosodic; no language parsing
No wake word. Language-agnostic at the acoustic level.
How it integrates
Where SAA sits in your drive-thru stack.
SAA sits between your lane mic and ASR. It returns an addressee decision in real time, before speech-to-text runs.
Signal in
Audio from your lane mic streams to SAA before any ASR or ordering model touches it. No change to your speaker post or menu board.
Addressee decision
SAA reads acoustic and prosodic signals to decide, before ASR, whether speech is addressed to the device or to someone else in the car.
Route only addressed audio
Only addressed audio reaches your ASR and ordering model. Back seat, radio, next lane, and crew crosstalk are suppressed.
Real-time on a live lane
SAA runs as a hosted GCP service, licensed per device. It fits inside your live lane latency budget; a DPA is available.
architecture
Pre-ASR, hosted, behind any model
Client SDKs in JavaScript/TypeScript and Python stream lane audio to server.attentionlabs.ai and receive the addressee decision before transcription begins.
SAA is model-agnostic. It works in front of any ASR or ordering model you already license, with no hardware changes.
access
Eval tier is free. Licensing is per device.
Request an eval to run SAA on audio from your own lanes.
Request an evalWhy existing tools miss it
Noise cancellation and VAD do not decide who is addressing the lane.
Each layer below solves a real problem, but none decides whether the ordering device is being addressed. That decision has to happen before ASR.
| Layer | What it answers | When it runs | Decides who is addressed? |
|---|---|---|---|
| Noise cancellation and AEC | Is the audio cleaner? | Pre-ASR | No |
| Voice activity detection (VAD) | Is someone speaking? | Pre-ASR | No |
| Speaker diarization | Who spoke, after the fact? | Post-ASR | No |
| Wake word | Was a trigger phrase said? | Pre-ASR | Phrase only |
| SAA addressee detection | Is the lane being addressed, right now? | Pre-ASR, real time | Yes |
Drive-thru and QSR voice platform teams who need the right person served.
Common questions
What QSR and platform teams ask before integrating.
What causes drive-thru voice AI to take orders from the wrong person?
The lane mic captures every voice in range: the driver, a back-seat passenger, the radio, the next lane, and crew talking behind the counter.
Without a layer that decides who is addressing the ordering device, all of it reaches your ASR and ordering model as potential order speech.
The recognizer isn't failing. It has no signal for who the speech was meant for. SAA adds that decision before transcription runs.
Does SAA replace our ASR, ordering model, or POS?
No. SAA is a pre-ASR layer in front of the stack you already run. It decides whether speech is addressed to the lane and forwards only that audio onward.
You keep your existing ASR, your ordering model, and your POS. SAA is model-agnostic and adds one decision upstream of the rest of the pipeline.
Where does SAA sit in the ordering pipeline?
Between the lane mic and your ASR. Raw audio streams to the hosted SAA service, which returns an addressee decision in real time, before STT.
Everything else is suppressed. When SAA is uncertain it fails closed and forwards nothing.
What is the latency on a live lane?
The routing decision returns before your ASR would have started, so it fits inside a live lane latency budget without slowing throughput.
SAA requires no GPU and no wake word at the speaker post.
How do we evaluate SAA on our own lane audio?
The evaluation tier is free.
Request an eval and we provide API keys, an eval kit, and the client SDK to test SAA on audio from your own lanes.
How is SAA licensed, and what does it cost?
SAA is licensed per device, not by usage or call volume. The evaluation tier is free.
Request an eval to run SAA on your own lane audio, or book a call to scope a deployment and discuss commercial terms.
How accurate is SAA on drive-thru lane audio?
SAA makes the addressee decision reliably on audio alone, across held-out multi-party sessions, in real time before STT (arXiv:2604.08412).
This is the conservative floor for open-mic environments such as a drive-thru lane with no video stream available.
Two caveats apply: it fails closed under distribution shift, and cross-lingual recall is a known limitation under active work.
How much voice-AI spend does a single drive-thru lane waste on non-actionable audio?
By bottom-up estimate, a single drive-thru location spends roughly $25,000 per year on ASR and model inference for audio never addressed to the ordering device.
That includes back-seat passengers, car radio, next-lane voices, and crew crosstalk, an estimated 86% non-actionable audio rate applied to per-minute inference pricing.
This is labeled ESTIMATE and varies by lane volume, ASR pricing, and session length. SAA routes only addressed audio to the stack, reducing inference on non-actionable audio.
Get in touch
Book a call or request an eval.
Tell us about your drive-thru voice stack and the lanes you run. We will follow up with next steps for an evaluation on audio from your own environment. The eval tier is free; licensing is per device.
Or try the Cloud SDK to get a feel.
Last updated June 29, 2026