Real-time addressee detection
stop your voice ai from answering everyone in the room.
Capabilities
Real-time Engagement Control System.
route
Addressee routing
Tags each utterance device-directed, human-directed, or ambient, before any ASR call runs.
attend
Engagement gating
Holds anything below engagement threshold. Ambiguous frames avoided.
stream
PCM handoff
On a positive gate, emits a turnReady event with the PCM16 segment, ready for any downstream model.
How it works
Audio in, events out, anywhere.
Connect the SDK
Reach the hosted Selective Auditory Attention (SAA) cloud SDK with your API key.
Send audio or video
Feed PCM16 audio, or add video for richer context. Either modality, anywhere.
Receive real-time events
Typed events per frame: speaker, VAD, and gate status, each scored and timestamped.
Forward only what matters
On a positive gate, turnReady carries the addressed segment; everything else is suppressed.
The stack
Where SAA sits in the voice stack.
Every voice product runs the same pipeline. One layer has been missing: whether the expensive rest should run at all.
-
01
Audio capture
The raw microphone stream from the room.
-
02
Voice activity detection
Is anyone speaking at all?
-
03
Wake word / trigger
Was a keyword spoken? Often skipped on always-on devices.
-
04
Addressee detection
Is this speech directed at the device, or at someone else? Routes only addressed audio onward, before ASR runs. Learn more about addressee detection.
the layer we build -
05
Speech recognition
Turning speech into text.
-
06
Language model
Understanding intent and forming a response.
-
07
Speech synthesis
Speaking the response back.
Two ways to deploy
Run SAA on-device or on the cloud.
Try it now
Cloud SDK
The low-friction way in.
- Hosted, pre-ASR engagement-control API
- JavaScript and Python SDKs; streaming and REST
- Same models and events as on-device
- Decision in about 150ms end to end, before ASR
- Per-utterance output: device-directed + confidence
- Adjustable decision threshold
- No wake word
- Start free, no card; usage-based after
- Engineering support on Slack
npm install @attenlabs/saa-js
# or
pip install attenlabs-saaEnterprise
Per-device license
Built for regulated and OEM hardware deployments.
- On-device inference; raw audio never leaves device
- ONNX runtime, CPU-only; no GPU
- Under 9ms inference, before ASR
- Per-utterance output: device-directed + confidence
- Audio, or audio and video fusion
- Acoustic and prosodic signals, designed to be language-agnostic
- Fails closed under uncertainty
- No wake word, no enrollment, no speaker ID
- Adjustable decision threshold via SDK
- Per-device license; custom acoustic profiles
- Per device, volume-based licensing
Hardware and software teams shipping addressee detection before ASR runs.
FAQ
Does SAA run on-device or in the cloud?
Cloud by default: audio streams to our hosted endpoint and the gate decision returns before ASR runs. On-device ships only as a separate enterprise license.
What does the layer actually receive?
PCM16 audio frames, plus camera frames if you add video. Nothing is retained after inference; see our privacy policy and DPA.
Can I use this with my existing voice AI stack?
Yes, it is model-agnostic. turnReady hands the addressed segment to OpenAI, Gemini, Pipecat, LiveKit, or your own model.
What modalities does the layer support?
Audio alone, or audio plus video for added spatial context. Both share the same event schema in the cloud SDK.
Can I evaluate with my own audio before committing?
Yes. New cloud accounts start free, no card required, with usage-based pricing. For an enterprise per-device license, book a call.
How do I ship this in hardware?
Through an enterprise per-device license, with custom acoustic profiles arranged on a call.
What languages are supported?
English is the primary evaluation set; multi-language support is underway. Contact us about your requirements.
Is this just voice activity detection (VAD)?
No. VAD only detects that someone is speaking, not whether they are addressing the device. SAA runs after VAD and answers that.
Is this speaker identification or diarization?
No. Speaker ID answers who is speaking; SAA answers who is being spoken to, with no enrollment required.
Does this add latency?
Yes, one hosted round-trip of about 150ms end to end before ASR runs; inference itself is under 9ms. In exchange, ASR, the LLM, and TTS never fire on audio that was never addressed to the agent.
Get in touch