The Engagement-Control Layer for Voice AI
The pre-ASR layer that decides who, if anyone, is addressing the device.
What engagement control is
Neither the ASR nor the model asks who this is for.
Every voice AI pipeline has an ASR and a model. Neither can answer the prior question: was this utterance addressed to the device.
Without that answer, every voice in range gets transcribed, background noise included. STT, LLM, and TTS costs follow all of it.
the gap
ASR and LLMs do not know who is speaking to them
An ASR transcribes every voice that reaches it. An LLM acts on whatever the transcript says. Neither carries a signal about who the speech was addressed to.
the fix
Engagement control answers the prior question
Before the ASR sees any audio, the engagement-control layer decides: is this utterance addressed to the device. Only addressed audio moves forward.
Non-addressed speech is dropped. When uncertain, it fails closed and forwards nothing.
Where it runs
Pre-ASR, hosted, on any device, behind any model.
Selective Auditory Attention (SAA) sits between the microphone and the ASR. It receives raw audio and returns a routing decision before transcription begins. No ASR swap, no GPU on customer hardware, no wake word.
position in the stack
Pre-ASR
Operates in the audio layer where acoustic addressee cues are still present.
deployment
Hosted on GCP
SAA is a hosted cloud service. Client SDKs in JavaScript/TypeScript and Python call server.attentionlabs.ai.
No model runs on customer hardware. A Data Processing Agreement is available for regulated use cases.
compatibility
Behind any model or ASR
SAA is model-agnostic and ASR-agnostic. Drop it in front of your current stack: no pipeline changes.
Where it sits in the stack
Underneath the models and the ASR, above the microphone.
The diagram below shows where the addressee decision happens in a complete voice AI system, and what becomes of speech that is not addressed to the device.
Performance
High accuracy on the addressee decision, audio-only or with audio-video fusion.
Measured on held-out multi-party sessions, published at arXiv:2604.08412. The system fails closed under distribution shift.
coverage
One decision, with or without a camera
The routing decision does not depend on video being available.
latency
Real-time decision, before STT
No GPU required to run the decision.
signal model
Acoustic; no language parsing
Language-agnostic at the acoustic level. Cross-lingual recall parity is a known open item under active work.
Why existing layers do not solve it
VAD, wake words, and diarization do not decide who is addressing the device.
Each layer below solves a real problem. None answers the prior question: is this utterance, right now, addressed to the device.
| 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 |
| Wake word detection | Was the trigger phrase said? | Pre-ASR | Trigger phrase only |
| Speaker diarization | Who spoke, and when? | Post-ASR | No |
| Intent classification / NLU | What did the transcript mean? | Post-ASR | No |
| SAA engagement control | Is the device being addressed, right now? | Pre-ASR, in real time | Yes |
How it integrates
One SDK stage in front of the ASR you already run.
SAA integrates as a single pre-ASR stage. Only addressed audio is forwarded downstream; your ASR, model, and hardware stay unchanged.
Audio in
Raw audio from your microphone streams to the SAA hosted service. Client SDKs in JavaScript/TypeScript and Python connect to server.attentionlabs.ai. No change to your microphone or downstream pipeline.
Addressee decision, pre-ASR
SAA analyzes acoustic signals in real time, before speech-to-text: is this utterance addressed to the device, or to someone else in the room. The decision uses signal that is present in the audio but never reaches a transcript.
Route only addressed audio
Only addressed audio reaches your ASR. Bystander speech, background conversation, and crosstalk are dropped at the gate. When uncertain, the failure mode is silence, not a false action.
Your ASR and model run on clean input
Your ASR and model run on addressed audio only. The layer is invisible to the rest of the stack: one decision upstream, nothing else changes. A DPA is available; licensing is per device.
access
Eval tier is free. Licensing is per device.
Request an evalOr try the Cloud SDK to get a feel.
Every voice AI system answers speech. Engagement control is the layer that decides which speech to answer, before the model runs.
Common questions
What builders and platform teams ask before integrating.
What is engagement control?
Engagement control decides whether the current utterance is addressed to the device, or whether it should stay silent.
It runs before transcription, underneath the ASR and the model.
SAA is the engagement-control layer built by attention labs. It returns a routing decision in real time, before speech-to-text, and fails closed when uncertain.
How is the engagement-control layer different from ASR or an LLM?
ASR transcribes whatever speech it receives, with no signal about who it was addressed to. An LLM parses intent after that addressing decision was already missed.
The engagement-control layer sits before both. Only addressed audio is forwarded; ambient speech never reaches the ASR or LLM.
Why can't the model or ASR vendor build the engagement-control layer?
Addressee cues live in the acoustic signal and are stripped out the moment audio becomes text.
The decision has to happen before transcription, using signal the transcript never carries.
Does SAA work behind any model or ASR?
Yes. SAA is model-agnostic and ASR-agnostic. It forwards only addressed audio to whatever ASR and model you already run, no swap required.
Client SDKs in JavaScript/TypeScript and Python connect to server.attentionlabs.ai.
What does the engagement-control layer ship as?
SAA ships as a hosted cloud service on GCP. Client SDKs in JavaScript/TypeScript and Python stream audio to server.attentionlabs.ai for a real-time routing decision.
No model runs on customer hardware. No wake word. No GPU. Eval tier is free; commercial licensing is per device.
How is engagement control different from voice activity detection (VAD)?
VAD detects whether someone is speaking; it does not know who they are speaking to.
Engagement control runs after VAD, on confirmed speech, and decides whether it is addressed to the device.
Only addressed speech is forwarded to the ASR. VAD is pre-ASR but not addressee-aware; the engagement-control layer is both.
Get in touch
Request an eval or book a call.
Describe your voice AI stack. We will follow up with eval access to run SAA on audio from your own environment. Eval tier is free; licensing is per device.
Last updated June 30, 2026