attention labs
Attention Labs provides a real-time, pre-ASR addressee detection layer called Selective Auditory Attention (SAA) that helps voice AI agents distinguish speech directed at them from ambient conversation. The system works without wake words, uses audio and video signals to infer engagement intent, and integrates with any voice stack via cloud SDK or on-device deployment. SAA achieves F1 0.95 with audio-video fusion and F1 0.86 audio-only on held-out test sets.
- Artificial Intelligence
- Developer Tools
- AI Agents
- Software Only
Funding
Funding not disclosed
Founders
Founder details are not available yet.
Product
Problem
Modern voice agents hear everyone in a room but cannot distinguish which speech is directed at them versus ambient conversation. This leads to false triggers, disrupted interactions, and poor user experiences in multi-party environments such as homes, offices, robotics, and spatial computing settings.
Solution
Attention Labs provides a real-time, pre-ASR addressee detection layer called Selective Auditory Attention (SAA). SAA runs before speech-to-text processing, determining whether spoken words are intended for the AI agent and filtering out side conversations and ambient noise. The system works without wake words, using audio and video signals to infer engagement intent, and integrates with any voice stack through a cloud SDK or on-device deployment. It is designed to fail closed when uncertain, suppressing responses when engagement confidence is low, thereby reducing false triggers in multi-party environments.
Target Audience
Primary customers are developers and product teams building voice AI agents for robots, XR/spatial computing platforms, smart devices, and desktop applications that operate in multi-party acoustic environments.
Features
- Real-time addressee detection that operates before ASR, routing only device-directed speech downstream
- No wake word requirement, enabling natural conversational interaction without hotword triggers
- Audio-video fusion architecture achieving F1 0.95, with audio-only F1 at 0.86 on held-out test sets
- Cloud SDK available via pip (`attenlabs-saa`) and npm (`@attenlabs/saa-js`) plus on-device deployment options
- Fail-closed design that suppresses responses when engagement confidence is low, reducing false triggers
- Demonstrated support for robotics (Reachy Mini), XR spatial computing multi-agent environments, and multi-speaker PC scenarios