The attention gate for real-world voice AI
Finally, voice AI that knows when to be quiet.
Build voice agents that know when to speak, and when to stay quiet.
drop it into any voice platform in just a few lines
# pip install attenlabs-saa
from saa import AttentionClient
saa = AttentionClient(
token=os.environ["SAA_API_KEY"],
initial_threshold=0.7,
)
saa.start()// npm install @attenlabs/saa-js
import { AttentionClient } from "@attenlabs/saa-js";
const saa = new AttentionClient({
token: "YOUR_SAA_API_KEY",
initialThreshold: 0.7,
});
await saa.start();
Modern voice agents need selective attention. Not better activation.
Field demonstrations
A robot ignores ambient speech and answers only when it is addressed. No wake word.
Several agents in one room, each knowing who is addressing it. Automatic turn-taking, no triggers.
Many speakers at once. The voice directed at the device is isolated, everything else suppressed.
The real world
Your voice agent hears everyone in the room.
It cannot tell who is talking to it.
without selective attention
answers every voice in the room
with selective attention
answers only the voice meant for it
Selective Auditory Attention (SAA)
Our first system, selective auditory attention (SAA), knows which speech is directed at the device, and plugs in before any model, on any device.
- who is being addressed,
- when models should speak,
- when they should stay silent.
the product
The engagement control layer for voice AI
Real-time, pre-ASR addressee detection for voice agents.
Plug in a Cloud SDK or embed on-device, native to any voice stack.
Cloud SDK
hosted, every platform
One SDK, with built-in integrations for LiveKit, Pipecat, Twilio, and OpenAI Realtime.
Embedded
on-device and offline
Runs natively on the device, no cloud round trip required.
- Runs before speech-to-text
- Fits any voice stack
- Answers you, ignores the room
Millions of years of evolution gave us selective hearing. And we've finally decoded it for machines.
Frequently asked questions
What is addressee detection?
Addressee detection determines whether speech in a multi-party environment is directed at a specific device or agent, as opposed to another person in the room. SAA by attention labs performs addressee detection in real time, before audio reaches your ASR or LLM.
How is SAA different from voice activity detection (VAD)?
VAD detects whether someone is speaking. SAA (Selective Auditory Attention) determines who is being spoken to. SAA routes only speech addressed to the device downstream, eliminating false triggers from side conversations, background voices, and ambient speech.
Does SAA require a wake word?
No. SAA requires no wake word. It uses audio and video signals to infer engagement intent, allowing natural conversation without hotword triggers.
How fast is SAA?
SAA runs inference in under 9ms; hosted end to end, the addressee decision returns in about 150ms, before ASR, well inside a real-time voice latency budget with no perceptible delay. See technology overview.
How accurate is SAA?
In our published evaluation, SAA’s audio-visual model reaches 95% F1. Accuracy varies with model size and the signals available. Lighter embedded models land around 90%, and the larger audio-visual models reach up to 95%, depending on whether they use audio, vision, or both. It runs in real time before speech-to-text and is designed to fail closed when uncertain. See the research page for details.
Is my audio stored?
No. attention labs runs the gate on a hosted endpoint and only forwards the speech addressed to your device; everything else is dropped. Raw audio and video are not kept beyond the moment they are processed, and are never used to train models. Encrypted in transit and at rest. More in our privacy policy, security page, and dpa.
Make your agent respond only when it's spoken to.
Building the engagement layer for voice AI