Intent-of-address gating for meeting and room-audio AI

Stop Side Conversations From Polluting Your Meeting Summary

Selective Auditory Attention (SAA) gates each turn before it reaches your meeting AI, so only addressed speech makes the recap.

Two participants hold an aside. The meeting AI writes it up as an action item.

In a many-to-many meeting, people address each other as often as they address the assistant. Nothing in a typical stack decides which is which before the assistant acts.

SAA (Selective Auditory Attention) closes that gap: it decides intent of address before the assistant acts, so side conversations stay out of the summary.

How SAA gates intent of address in a meeting room A conference room birds-eye view. Two participants address the meeting assistant (lime arrows forwarded). Two hold a side conversation (grey dashed line, gated out). A summary panel on the right shows only the addressed turns included; the side conversation is excluded. Meeting AI SAA gate A Participant A addressed B Participant B addressed C Participant C side conv. D Participant D side conv. aside SUMMARY OUTPUT included Participant A forwarded to assistant Participant B forwarded to assistant excluded by SAA gate C + D: side conversation gated out, not included Addressed to meeting assistant (forwarded) Side conversation (excluded from summary)
Lime marks turns addressed to the assistant; grey dashed marks the side conversation. SAA gates each turn before the assistant processes it.

false inclusion

Side conversations pollute the recap

A brief aside between two participants becomes a bullet point in the summary, with a follow-up nobody owns.

wrong attribution

Action items land under the wrong name

The system knows who spoke, not who they were speaking to, so follow-ups get assigned to people who were only talking to each other.

A hosted gate that decides intent of address before the assistant ever acts.

SAA is a hosted, pre-ASR layer that decides whether speech in a meeting room is addressed to the AI assistant or to another participant. It routes only addressed audio to the model stack you already run, fails closed when it is unsure, and needs no wake word and no GPU.

Published on multi-party sessions.

SAA reliably decides who is being addressed on held-out multi-party sessions, the same structure meeting rooms produce (arXiv:2604.08412).

Two caveats apply: it fails closed under distribution shift, and cross-lingual recall is a known limitation under active work.

input modality

Works with or without a camera

Uses the room's camera feed alongside audio when one is present, and falls back to audio alone when it is not.

latency

Decision in real time, before STT

The intent-of-address decision returns before ASR begins, inside a live meeting turn budget.

design principle

Fails closed by design

Ambiguous speech is not forwarded. Under uncertainty, SAA stays silent rather than risk including the wrong turn.

One stage before your ASR. No hardware change.

SAA sits between your room audio and your ASR, deciding intent of address before speech-to-text. Only addressed turns reach your meeting-intelligence model.

Your transcription stack, your ASR vendor, and your room hardware stay exactly as they are.

How SAA routes only addressed audio in a meeting room Room microphone or meeting-agent audio streams into the hosted SAA service. SAA returns an addressee decision in real time, before ASR. Audio addressed to the assistant is routed to the existing ASR and meeting-intelligence model. Side conversations, ambient room noise, and anything uncertain are not forwarded; the system fails closed. Room audio mic or agent track all participants SAA addressee decision real time, pre-ASR Addressed to ASR + meeting AI Not addressed side talk, held at gate
SAA sits before ASR. Addressed speech is forwarded to your existing ASR and meeting-intelligence stack; side conversations, ambient noise, and anything uncertain stop at the gate.
01

Stream audio to SAA

Audio from the room microphone or meeting-agent audio track streams to the hosted SAA service through a thin client SDK in JavaScript/TypeScript or Python.

No model runs on your hardware. No wake word. No GPU required.

02

Intent-of-address decision

SAA analyzes acoustic and prosodic patterns before ASR runs, without parsing language content. It is language-agnostic at the acoustic level.

Before speech-to-text, it returns a decision: was this turn addressed to the assistant, or was it an exchange between participants?

03

Forward only addressed turns

Turns addressed to the assistant are forwarded to your ASR and meeting-intelligence model.

Participant-to-participant speech and ambient room noise are suppressed. Under uncertainty, SAA forwards nothing.

04

Hosted, per-device, DPA available

SAA runs as a hosted service on GCP at server.attentionlabs.ai, model-agnostic in front of any ASR and meeting-intelligence stack you already license.

Licensing is per device. A DPA is available on request.

integration path

Software layer only, no hardware change

One thin SDK stage in front of your existing pipeline. Your room hardware, microphone array, and ASR vendor stay unchanged.

access

Eval tier is free. Licensing is per device.

Or try the Cloud SDK first to get a feel for the integration.

Request an eval

Diarization knows who spoke. It does not know whether they were talking to you.

Every layer your meeting stack runs today solves a real problem. VAD catches speech. Diarization labels the speaker. Beamforming cleans the signal.

None of them decides whether the assistant is the one being addressed. That is why side conversations still reach the summary.

Layer What it answers When it runs Decides intent of address?
Voice activity detection (VAD) Is someone speaking? Pre-ASR No
Beamforming and AEC Is the audio cleaner or spatially filtered? Pre-ASR No
Speaker diarization Who spoke, and when? Typically post-ASR No
Speaker ID Is this a known voice? Typically post-ASR No
Wake word Was a trigger phrase said? Pre-ASR Phrase only, not intent
SAA addressee detection Is the assistant being addressed right now? Pre-ASR, before STT Yes
Voice activity detection (VAD)
What it answersIs someone speaking?
When it runsPre-ASR
Decides intent of address?No
Beamforming and AEC
What it answersIs the audio cleaner or spatially filtered?
When it runsPre-ASR
Decides intent of address?No
Speaker diarization
What it answersWho spoke, and when?
When it runsTypically post-ASR
Decides intent of address?No
Speaker ID
What it answersIs this a known voice?
When it runsTypically post-ASR
Decides intent of address?No
Wake word
What it answersWas a trigger phrase said?
When it runsPre-ASR
Decides intent of address?Phrase only, not intent
SAA addressee detection
What it answersIs the assistant being addressed right now?
When it runsPre-ASR, before STT
Decides intent of address?Yes
Built for

Meeting-agent and room-audio teams whose product promise is a clean, trustworthy recap.

What meeting-platform and hardware teams ask before integrating.

Why do side conversations end up in my meeting summary?

VAD and diarization tell your meeting AI that someone spoke and who spoke, not whether they were addressing the assistant. When two people hold a side conversation in the corner of the room, the assistant has no principled way to exclude it, so the aside lands in the summary anyway.

Is diarization enough to fix attribution?

No. Diarization tells you who spoke, usually after transcription, not whether they were addressing the assistant. SAA adds that intent-of-address decision upstream, before the assistant acts on any turn.

How does SAA decide intent of address in a meeting room?

SAA analyzes acoustic and prosodic signals in the audio stream, no wake word, no language parsing, no GPU required. It decides in real time, before speech-to-text, and fails closed when uncertain, meaning it does not forward the audio or trigger a summary entry.

Does SAA integrate at the hardware level or through the meeting agent?

Through the meeting-agent or room-audio software layer, not hardware. A thin SDK gates which turns reach the assistant, no firmware change needed, and SAA is model-agnostic in front of whichever ASR and meeting-intelligence stack you already run.

How do we evaluate SAA on our own meeting audio?

The evaluation tier is free. Request an eval and we provide API keys, an eval kit, and the client SDK so you can run SAA against audio representative of your own meeting rooms.

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 it on your own meeting audio, or book a call for commercial terms.

A clean summary starts with knowing who was talking to the assistant.

Tell us about your meeting platform or room-audio hardware and what you are trying to improve. We will follow up with next steps for a free evaluation on audio from your own environment.

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