SAA decides whether speech was meant for a device before it reaches the voice AI stack, so agents respond only when addressed.

SAN FRANCISCO, CA, UNITED STATES, June 24, 2026 /EINPresswire.com/ -- Attention Labs today launched SAA (Selective Auditory Attention), a hosted service that determines whether speech was addressed to a device before the audio reaches the speech-to-text, language model, and text-to-speech stack. The company describes the layer, which it calls engagement control, as a missing infrastructure component for voice AI deployed in shared spaces.

Voice agents operate reliably with a single user in a quiet setting, but they can respond to speech that was not directed at them in shared environments such as drive-thru lanes, vehicles, kiosks, and meeting rooms. Existing components address related problems: voice activity detection identifies that speech is present, diarization labels who spoke, and speech recognition converts speech to text. None determines whether a given utterance was addressed to the device.

SAA runs before speech recognition and returns one addressee decision per utterance. Speech directed at the device continues to the downstream stack, and other speech is held back. The service is model-agnostic, requires no wake word, and is designed to fail closed, remaining silent when confidence is low rather than responding incorrectly. SAA performs addressee detection and is distinct from transcription, diarization, voice activity detection, wake-word systems, and audio cleanup.

In the company's launch demonstration, two laptops in the same room run the same Grok voice model. The laptop without SAA responds to nearby speech regardless of the intended listener, while the laptop running SAA responds only when it is addressed.

“A voice agent can understand every word and still not know whether those words were meant for it,” said David J. Kim, co-founder and chief executive officer of Attention Labs. “SAA makes that decision before the audio reaches the rest of the stack, so an agent responds when it is spoken to and stays silent when it is not.”

SAA is available as a hosted classifier service accessed through JavaScript and Python SDKs, with integration clients and runnable examples for LiveKit, Pipecat on Daily, ElevenLabs Conversational AI, Twilio Media Streams, and OpenAI Realtime. These reflect framework compatibility and are not partnerships or endorsements. Enterprise and OEM customers can license local or embedded deployment. The underlying model and weights remain proprietary to Attention Labs. A research paper, titled Selective Attention System, documents the method. The published evaluation is primarily in English, and cross-lingual performance is an active area of work.

Attention Labs holds a granted United States patent on its core arbitration mechanism, within a portfolio of more than 45 filed and issued patents. SAA received CES 2026 Picks recognition from TWICE, Residential Systems, and TechRadar Pro. The company is backed by Techstars, Motivate Ventures, and Andreessen Horowitz (a16z), and is based in San Francisco and Toronto.

More information is available at https://attentionlabs.ai/.

About Attention Labs
Attention Labs builds engagement control infrastructure for voice AI. Its first product, SAA (Selective Auditory Attention), is a hosted addressee detection service that determines whether speech was directed at a voice agent before it reaches the speech-to-text, language model, or text-to-speech stack, so that systems deployed in shared spaces respond when addressed and stay silent when they are not. The company holds a granted United States patent on its arbitration mechanism, within a portfolio of more than 45 filed and issued patents, is backed by Techstars, Motivate Ventures, and Andreessen Horowitz (a16z), and is based in San Francisco and Toronto.

Press Office
Attention Labs
press@attentionlabs.ai

Legal Disclaimer:

EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.