Pinned Tweet
AI that knows the room: stays silent, speaks on cue, and routes to the right model (local or cloud).
No wake words. No cameras. On-device.
Demo ↓ 🔊
We just gave every voice model selective hearing 👂⚡️
SAA lets any device tell whether speech is meant for it or just happening nearby, even with multiple talkers. No wake word required.
JS/Python + LiveKit, Pipecat, ElevenLabs & Twilio:
github.com/attenlabs/saa-sdk
Starting attention labs taught me that the hardest problems are the ones you can’t hear.
In the early days we tackled the infamous hybrid meeting howling (echo) problem.
Cracked that in 2022.
But the real challenge we uncovered was teaching machines to listen like humans, to focus on what matters and ignore the rest.
So we turned to neuroscience. Brains filter conversations in milliseconds. One overlap can derail understanding, but the smallest improvement in perception makes the whole exchange flow. That became our blueprint.
Tackling invisible problems demanded patience, trust, and curiosity. Science, engineering, and humility became our principles.
Today, our selective attention AI helps people communicate more clearly, focus better, and experience conversations naturally, even in noise. The hardest problems are invisible, but solving them changes everything.
2
17
735
We gave AI the ability to know when to speak.
No wakewords. No cameras. 100% local.
Two voice agents joined a meeting on our laptops, catching questions, waiting their turn, and answering in real time.
Powered by our selective auditory attention engine:
🎯 97% accuracy on speaker detection
⏱️ Speaks at the right time, ignores false triggers
📦 400k params, fully on-device
🔌 Works with any LLM
This is the next wave of natural human-AI collaboration.
3
13
775
Geoff Hinton came to Toronto in ’87 when no one else funded AI, found an open haven at UofT.
From backprop → Google Brain → OpenAI’s Ilya → Cohere & Vector Institute… Toronto’s been at the heart of deep learning.
If 🇨🇦 doubles down on funding, immigration & incentives, the next Hinton/Cohere will stay.
We’re hiring here in Toronto!
Join us in building the auditory attention layer for AI
meta’s demo didn’t fail because of wifi or a weak llm
it failed because the input chain broke
wifi always collapses with big crowds
that’s why we run local on-device and make the listening layer robust
until ai can really hear even the smartest models will keep breaking in real-world conversatinoal scenarios..
bottleneck isn’t bigger brains it’s better ears.
3
16
668
