Pre-ASR addressee gate for voice agent pipelines

Your pipeline runs ASR on everything.

SAA (Selective Auditory Attention) routes only the audio actually addressed to your agent.

Your pipeline can't tell who speech was for. It processes all of it anyway.

VAD fires on any sound above a noise floor. Nothing in your current stack decides whether that speech was actually addressed to the agent.

Mid-sentence, the agent hears background noise, stops, and reruns ASR on audio the caller never directed at it. The caller assumes it's broken.

VAD fires on all voices; no addressee signal reaches the agent Three audio sources reach the VAD: the caller's addressed speech, background speech in the caller's room, and a second speaker. VAD passes all three because it only detects that sound is present, not who is being addressed. The agent then stops, resumes awkwardly, or acts on non-caller speech. Caller (addressed) Background speech Second speaker Voice activity detection fires on all three sources Agent receives all three false stops / false triggers / wrong turns
Lime marks the actual caller. VAD passes all three sources because it detects sound, not who is being addressed. The agent then reacts to voices it should have ignored.

false barge-in

Background speech triggers a false stop and burns ASR budget for nothing

The agent interrupts itself and reruns ASR on speech meant for someone else in the room. The caller hears an awkward restart, and your platform absorbed the cost.

missed real interruption

Real corrections get lost while the agent waits out background noise

The caller's actual correction goes unregistered until noise drops below the VAD threshold. The turn is lost, the call re-routes, and handle time goes up.

One pre-ASR gate. Route only addressed audio downstream.

SAA is a hosted gate that decides, before ASR runs, whether speech was addressed to the agent.

Addressed audio forwards to your ASR and LLM exactly as today. Everything else is suppressed before ASR runs.

Built on our published technical paper. Quantified on real multi-speaker audio.

SAA reliably detects addressed speech, using audio alone or audio-and-video fusion, on held-out multi-party sessions (arXiv:2604.08412).

Two caveats: it fails closed under distribution shift, and cross-lingual recall is a known limitation. Run the eval on your own call audio to see how it performs on yours.

performance

Reliable on audio and video fusion

Held-out multi-party sessions, published at arXiv:2604.08412.

performance

Reliable on audio-only

Real multi-speaker audio, no video stream required. Same published results.

latency

Real-time routing decision

Returns before your ASR would have started, no GPU required, and fits inside a real-time agent latency budget on any transport.

design principle

Fails closed by design

The default under uncertainty is silence, not a false barge-in or a false stop.

coverage

Multi-speaker environments

Multi-party calls, conference bridges, and noisy caller environments with concurrent speakers.

signal model

Acoustic; no wake word

Learned acoustic signals at the audio layer. No trigger phrase required.

One import. Sits in front of your VAD. Zero pipeline changes.

The addressee decision returns before your ASR starts processing.

No ASR swap, no LLM changes, no rewiring. Works over WebRTC, SIP, and raw audio transports.

SAA as a drop-in pre-ASR stage in a voice agent pipeline The voice agent pipeline runs left to right: microphone or transport audio enters, then streams to the SAA pre-ASR gate which decides in real time whether speech is addressed. Addressed audio passes to ASR, then to the LLM, then to TTS output. Non-addressed speech and uncertain audio are suppressed at the SAA gate, never reaching ASR or the LLM. Mic / transport all audio in SAA addressee gate pre-ASR, real time ASR your existing LLM your existing TTS output Not addressed / uncertain suppressed, not forwarded one SDK import, no pipeline changes
SAA sits between audio input and ASR. Addressed caller speech (lime arrow) passes through. Background speech, second speakers, and anything uncertain are suppressed before ASR runs. Your ASR, LLM, and TTS are unchanged.
01

Stream audio to the hosted SAA service

JavaScript/TypeScript and Python client SDKs stream audio from your telephony transport or mic to server.attentionlabs.ai.

No model runs on your hardware.

02

Receive the addressee decision in real time

SAA analyzes acoustic signals and returns a routing decision in real time: addressed, or not.

No content is parsed, no transcript produced.

03

Forward only addressed audio to your ASR

Addressed audio forwards to your existing ASR and LLM, exactly as before.

Background speech, second speakers, and anything uncertain are suppressed. Fails closed: when unsure, SAA forwards nothing.

04

Works in front of any pipeline architecture

Inserts as one routing stage before your VAD, or directly after your audio source.

One SDK import, one forwarding condition.

architecture

Hosted, pre-ASR, behind any model

SAA runs as a hosted GCP service. Thin client SDKs handle the audio streaming. No model to deploy, no GPU to provision.

Licensed per device. A DPA is available for regulated deployments.

access

Free eval. Per-device licensing.

Request an eval and we provide API keys, an eval kit, and the client SDK to run SAA against your own call audio, or try the Cloud SDK first.

Request an eval

VAD, beamforming, diarization, and wake words each answer a real question. None answer this one.

Every layer below does its job correctly. None decide whether the speech reaching your pipeline was addressed to your agent. That decision has to happen before ASR runs, or every voice that clears the noise floor reaches the part of your pipeline that acts on it.

Layer What it answers When it runs Decides who is addressed?
Voice activity detection (VAD) Is someone speaking? Pre-ASR No
Beamforming and acoustic echo cancellation Is the audio cleaner / which direction? Pre-ASR No
Speaker diarization Who spoke, and when? Post-ASR No
Wake word Was a trigger phrase said? Pre-ASR Trigger phrase only
Speaker identification Is this a known voice? Post-ASR No
SAA addressee detection Is the agent being addressed, right now? Pre-ASR, real time Yes
Built for

Voice infrastructure teams done paying ASR and LLM costs on unaddressed audio.

What platform and infrastructure teams ask before integrating.

We already have barge-in and turn detection built in. What does SAA add?

VAD fires on any sound above a noise floor. SAA answers a different question: was this speech addressed to the agent, or is it background conversation, a second person, or hold-music bleed?

SAA sits upstream of your VAD as an addressee gate. It doesn't replace your turn-detection logic, and it doesn't touch your ASR, LLM, or TTS.

Does SAA replace our ASR, LLM, or TTS?

No. SAA is a pre-ASR gate in front of the pipeline you already run. It forwards only addressed audio to your ASR; your ASR, LLM, and TTS are unchanged. SAA is model-agnostic.

Where does SAA sit relative to our VAD?

SAA sits before your VAD, or can replace the VAD gate entirely. VAD detects that someone is speaking; SAA decides whether that speech was addressed to the agent.

Real interruptions pass through. Background speech and second speakers are suppressed. Uncertain audio fails closed.

How do I integrate SAA into my voice pipeline?

Client SDKs in JavaScript/TypeScript and Python stream audio to the hosted SAA service. The addressee decision returns in real time, and you forward only addressed audio to your ASR. One import, one forwarding condition.

What does the API return and what is the latency?

A routing decision: addressed (forward to ASR) or not addressed (suppress), returned before your ASR would have started. No GPU or wake word required.

How do I get started without a sales call?

Request an eval for API keys, an eval kit, and the client SDK to test on your own call audio, free. Or explore first with the Cloud SDK, no sales call required.

How is SAA licensed?

Per device, not by usage or call minutes. The evaluation tier is free. Request an eval to test on your own audio, or book a call to scope a deployment. A DPA is available on request.

Request an eval or book a call.

Tell us about your voice pipeline and the failure mode you want to close. We'll follow up with API keys, an eval kit, and the client SDK.

The eval tier is free. Licensing is per device.

We do not share your information. See our privacy policy.