Selective auditory attention is the human ability to lock onto the one voice speaking to you and tune out the rest of the room. Attention Labs builds that ability for machines. Its first system, SAA (Selective Auditory Attention), is a real-time attention layer for voice AI: it sits before speech-to-text and decides, per utterance, whether speech was actually addressed to the device, forwarding only what was meant for it and staying silent on the rest, with no wake word. The company calls the category attention arbitration, also described as addressee detection, device-directed speech detection, or engagement control: deciding who is being addressed, when an agent should speak, and when it should stay silent.
The SAA project README on GitHub: SAA makes one addressee decision per utterance, so only speech meant for the device reaches the STT, LLM and TTS, with no wake word.
For seventy years, the hard problem in machine listening has been the cocktail party problem: a room full of voices, and the need to attend to exactly one of them. Voice assistants mostly dodged it with a lexical crutch, the wake word. Say the magic phrase and the machine starts transcribing; otherwise it pretends not to hear. That works in a quiet kitchen. It falls apart the moment a device lives in a shared, noisy, multi-talker space and is expected to behave naturally, without a ritual phrase before every request. A voice agent in that setting answers everything it hears. The next breakthrough in voice AI may be less about hearing better and more about learning when not to listen.
Why the failure happens before transcription
Most of the voice stack is built to process audio, not to decide whether it should. Voice activity detection notices that someone is speaking. Diarization labels who spoke. Turn detection decides when a speaker finished. Speech-to-text transcribes. Every one of these assumes the audio was meant for the system. None of them answers the prior question: was this utterance addressed to the device at all? A bystander’s conversation passes voice activity detection cleanly and gets transcribed in full.
SAA answers that prior question. It classifies incoming audio in real time as device-directed, human-directed, or ambient, and forwards only the device-directed speech downstream, staying silent on the rest. That is a different job from the layers around it.
Where SAA sits in a voice pipeline: a pre-ASR addressee gate slotted between voice activity detection and speech-to-text, separate from noise suppression, diarization, end-of-turn detection, STT, LLM and TTS.
Each layer in the stack answers its own narrow question. Noise suppression cleans the signal. Voice activity detection asks whether anyone is speaking. Diarization labels who spoke. Turn detection estimates when a speaker finished. A wake word waits for a trigger phrase. None of them decides whether the speech was meant for the device. A denoiser such as Krisp and SAA are complementary rather than competing: the right order is denoiser, then the SAA gate, then speech-to-text, then the language model. Denoising makes the signal cleaner; SAA decides whether that signal should be acted on.
What it costs to listen to everything
When a system transcribes every utterance in earshot, it pays for every utterance in earshot. In a drive-thru lane, ambient engine idle, wind buffet, occasional radio bleed from the vehicle, and crosstalk from the adjacent ordering lane all hit the microphone at once. A conventional stack transcribes the customer’s order, the passenger’s side comment, and the neighboring car’s conversation alike, and downstream language-model inference runs on all of it. The same dynamic plays out in hotel lobby kiosks and airline check-in terminals, where always-on speech-to-text burns inference budget on staff chatter and bystander speech.
Placing the gate before speech-to-text means downstream transcription and language-model inference run only on addressed utterances. In shared-space environments with heavy background speech, that directly reduces per-session token and API spend. The economic argument and the accuracy argument point the same way.
On accuracy, Attention Labs cites published results of 0.95 F1 with audio-visual fusion and 0.86 F1 audio-only, in its paper “Selective Attention System: Device-Addressed Speech Detection” (arXiv:2604.08412). Two caveats travel with those numbers and should not be dropped. First, the audio-only 0.86 figure requires no camera, while the 0.95 figure depends on audio-visual fusion, so deployments that cannot add a camera should plan to the lower ceiling. Second, the model fails closed under distribution shift: when an utterance is ambiguous or out of distribution it is suppressed rather than passed, which is the conservative failure mode you want for routing. Cross-lingual recall is a known, data-bound limitation stated in the paper, with English as the primary language and broader coverage underway; non-English coverage should be validated rather than assumed. These are paper-scoped results, not deployment benchmarks for any specific room.
Why the gate runs before speech-to-text
The cleanest way to think about engagement control is as input authorization, in the same family as authentication and permissioning. Authentication decides who you are; permissioning decides what you may do; addressee detection decides whether this input should be acted on in the first place. Like its cousins, the gate belongs before the expensive work, not after it.
Illustrative shared-space demo footage, not a customer deployment. Shared rooms produce valid speech that may still be meant for someone else.
It is worth being precise about what this layer is not. SAA does not identify or authenticate the speaker; “was this order meant for the device” is a different question from “who is ordering,” and identity-gated workflows still need separate speaker-verification tooling. It does not transcribe, since it operates before speech-to-text and emits only a routing decision. It has no access to semantic content and is not a content-moderation layer. Knowing the boundaries is what makes the boundary it does hold credible.
Where this matters most
Engagement control earns its keep wherever a device lives in a shared physical space and bystander speech is the norm rather than the exception. Attention Labs’ public field demonstrations run on robots, including a Reachy Mini, in XR and spatial computing, and on personal computers. The same per-utterance gate generalizes to other shared-space settings the company describes as application areas it is exploring, not live customer deployments.
Illustrative robot demo, not a customer deployment. On an embodied machine, acting on the wrong speaker becomes part of the control path.
Robotics and embedded agents. On an embodied machine, acting on the wrong speaker is not a transcript error; it becomes part of the control path. On-device addressee gating for robots in shared spaces is licensed per device through enterprise and OEM terms.
Drive-thru and QSR voice ordering. A lane speaker hears engine idle, wind, radio bleed, and crosstalk from the next lane, while the customer is often half-turned toward a passenger. A gate that classifies whether the speaker was addressing the order point, before a single token reaches the transcription API, attacks the cross-lane false-trigger problem at the root. Under heavy noise the fails-closed behavior suppresses rather than guesses, which is the right default for an order.
Airport and airline self-service kiosks. A check-in or boarding kiosk sits in one of the hardest acoustic environments in commercial deployment: continuous PA announcements, crowd noise, and a passenger population spanning many languages in a single shift. A gate keyed to whether speech was directed at the kiosk, rather than to a keyword, targets the PA-triggered and bystander-triggered activations that push many kiosk voice stacks back to push-to-talk. Because airside connectivity is intermittent and passenger data is sensitive, the on-device path matters as much as the accuracy number, and the cross-lingual limit above is a real constraint to validate per language.
Hotel front-desk and concierge kiosks. A lobby kiosk faces a guest addressing the screen while a companion narrates aloud, or while staff conversation carries across an open atrium. The per-utterance gate is designed to activate on the guest speaking to it, not on the ambient conversation in the same space. Audio classified as non-addressed can be discarded at the edge before any cloud transcription occurs, which gives a property a concrete answer about what happens to overheard guest speech: it never enters the pipeline.
Why silence becomes a feature
A concierge kiosk that answers every guest who walks past, or responds to a television in the corner, is not merely wasting compute; it is failing at its job. The product quality users actually notice is restraint. An assistant that stays quiet when it was not being addressed reads as attentive, while one that interjects reads as broken. Reframing silence from a gap in coverage into a designed behavior is the conceptual shift engagement control makes routine, and it is what lets natural, wake-word-free interaction work in a room full of other people.
A model-agnostic layer
SAA is designed to be composable. It sits between the microphone and the speech-to-text engine and plugs into the voice infrastructure teams already use, including OpenAI Realtime, Gemini Live, ElevenLabs, Pipecat, and LiveKit. Only the addressed audio it forwards reaches the downstream speech-to-text engine, so transcription cost is spent on addressed speech alone. Device-directed speech detection has established prior art in consumer-assistant research, including work published by Apple’s machine-learning group, and arXiv:2604.08412 extends that line.
A frame from Attention Labs’ launch demonstration: two laptops in the same room run the same realtime voice model. The ungated agent (left) answers a remark addressed to another person, while the SAA-gated agent (right) recognizes it was not being spoken to and stays silent. Illustrative demo, not a customer deployment.
The SDKs are available in JavaScript (@attenlabs/saa-js) and Python (attenlabs-saa) on GitHub and connect to a proprietary hosted classifier; the hosted cloud API is the default entry point, with 500 free minutes to start and usage-based pricing after, no card required. On-device and embedded deployment, the path that matters for airside kiosks and robots, is licensed per device through enterprise and OEM terms, with royalty terms and custom acoustic profiles. The model is not open and is not freely self-hostable; buyers with strict data-residency or PII requirements engage on OEM terms, where audio need not be retained after the routing decision is made.
What to watch next
Attention Labs is backed by Andreessen Horowitz, Techstars, Motivate Ventures, Plug and Play, and Communitech, and SAA was named a CES 2026 TechRadar Pro Pick. The company frames the verticals above as application areas it is actively exploring rather than live deployments, and evaluates new ones through paid proof-of-concept engagements, where the open questions the published numbers cannot settle by themselves, real-environment performance under distribution shift, the end-to-end latency the per-utterance step adds, multi-device coordination in dense layouts, and non-English recall, are the right things to measure before any production commitment.
The Attention Labs team at CES 2026.
The larger bet reflects the company’s stated mission, to give people and AI a seat at the same table, since many of the most consequential decisions are still made synchronously, in groups. As voice agents move off the phone and into shared rooms, lanes, lobbies, and machines, the differentiator will not be how well a system transcribes. It will be whether the system knew, before it spent a token, that it was being spoken to at all.