Category definition · Engagement control

Addressee detection: the layer the voice stack was missing.

The pre-ASR layer that decides whether speech is addressed to your device.

What addressee detection is and what it is not.

Addressee detection classifies whether speech is directed at the device or at someone else, in real time.

It is a pre-ASR gate: raw audio is classified before transcription, and only addressed speech moves downstream.

Most speech near a voice AI is not meant for it: a drive-thru lane, an open office, an exam room, a home with several people. Without this gate, all of it reaches the ASR engine and the model.

The decision draws on acoustic signals present in raw audio and gone once transcribed. That is what makes it pre-ASR.

VAD detects speech. Wake words detect a phrase. Beamforming finds direction. Diarization labels the speaker later.

None of them answer is the device the intended recipient?

Voice AI hears everything in the room. Nothing decides who is being addressed.

A microphone in a real room is not a phone line. It captures bystanders, background media, and ambient noise alongside whoever is actually addressing the device.

Three audio sources reach one microphone; SAA forwards only the addressed one User A, who is addressing the device, has a lime-colored path through the SAA gate to the ASR. User B's side conversation and background ambient noise both reach the microphone as grey dashed lines but are blocked at the SAA gate and not forwarded. User A addressing device User B side conversation Background ambient noise MIC SAA before STT Addressed forwarded to ASR Not addressed blocked (fails closed)
Three sources reach the same mic. Selective Auditory Attention (SAA) forwards only the one addressed to the device. Uncertain cases fail closed.

environment mismatch

A single-user stack deployed in a multi-party world

VAD and wake words assume one user at a time. Voice AI now runs in lanes, lobbies, and vehicles with many speakers.

cost

Every unaddressed utterance generates real ASR and model spend

Voice agents pay for STT, LLM, and TTS on every second that enters the pipeline, addressed or not. SAA drops it first.

high-stakes settings

Wrong-person responses carry real consequences

A wrong drive-thru order costs margin. A wrong exam-room attribution is a documentation error. A false activation in a vehicle is a safety event.

A hosted pre-ASR gate. Additive to any existing stack. No ASR swap required.

SAA is a hosted, pre-ASR engagement-control layer. It streams audio, returns a routing decision before ASR runs, and forwards only addressed speech. Uncertain cases fail closed.

01

Signal in

Audio, and optional video, streams to the hosted SAA service through a thin JavaScript or Python client SDK.

No model runs on the device. No wake word. No GPU required.

02

Addressee decision

SAA analyzes acoustic patterns in the raw audio. It does not transcribe speech or parse language content.

Before speech-to-text runs, it returns a routing decision: addressed to the device, or to someone else in the room.

03

Route only addressed audio

Only audio addressed to the device reaches your existing ASR and model stack. Everything else is dropped.

When SAA is uncertain it fails closed and forwards nothing. The rest of the pipeline is unchanged.

architecture

Works in front of any ASR and model

SAA is model-agnostic. Drop it in front of any ASR and language model you already license.

No integration swap required. A DPA is available for regulated deployments.

try it first

The Cloud SDK is self-serve

JavaScript, Python, and REST, live now, so you can get a feel for the routing decision before requesting an eval.

Explore the Cloud SDK

The standard stack answers other questions, not who is being addressed.

These techniques get conflated because they all analyze audio in multi-speaker environments. None of them determine whether the device is being addressed. That is the specific job of addressee detection.

Technique Question answered Runs before ASR? Requires wake word? Addressee signal?
SAA addressee detection Is this speech directed at the device right now? Yes No Yes
Voice activity detection (VAD) Is someone speaking? Yes No No
Beamforming and AEC Is the signal clean or directional? Yes No No
Wake word detection Was the trigger phrase said? Yes Yes, it is the mechanism Phrase only
Speaker diarization Who spoke, and when? No, typically post-ASR No No

Measured performance, caveats included.

SAA reliably decides whether speech is addressed to the device, using audio-video fusion or audio alone (arXiv:2604.08412).

Two caveats apply: the system fails closed under distribution shift; cross-lingual recall is a known limitation under active work.

performance

Accurate on audio alone, or with video

The same addressee decision holds on audio only, or on audio and video fusion when a camera is available. No video stream is required.

latency

Real-time decision

Returned before ASR begins processing the utterance. No GPU required on the device side.

design principle

Fails closed by design

Ambiguous speech is suppressed: under uncertainty, the default is silence, not a false response.

A well-calibrated addressee detector would rather miss an intended utterance than act on an unintended one.

The missing layer

VAD tells you someone spoke. Addressee detection tells you who they were speaking to.

What OEMs and developers ask about the addressee layer.

What is addressee detection for voice AI?

Addressee detection is the pre-ASR classification of whether a spoken utterance is directed at the device or at another person. VAD, wake words, and diarization each answer something different; none of them determine whether the device is the intended recipient. Addressee detection is the decision that fills that gap. SAA performs it in real time, before ASR runs, and fails closed when uncertain.

How is addressee detection different from voice activity detection (VAD)?

VAD tells you that someone is speaking. It fires on background conversation, hold music, and bystanders equally. Addressee detection answers the question VAD cannot: is this speech directed at the device. SAA makes that call in real time, before ASR, and fails closed when uncertain.

How is addressee detection different from a wake word?

Wake words require a fixed trigger phrase before every interaction and still misfire when background speech contains the phrase. Addressee detection requires no trigger phrase, uses acoustic signals, needs no GPU, and fails closed when uncertain.

How is addressee detection different from speaker diarization?

Diarization tells you who spoke and when, after transcription. Addressee detection answers a different question: whether the utterance was directed at the device. Knowing the speaker label does not tell a system whether to act. SAA is complementary to diarization, not a replacement.

Why does addressee detection have to run before ASR?

The addressee signal lives in acoustic features present in raw audio, such as pitch, energy, and turn-taking cues. Once speech is transcribed that signal is gone. Running the gate pre-ASR is the only way to prevent unaddressed audio from entering the pipeline and generating spurious model calls.

Where is addressee detection used in production?

Any multi-party voice AI deployment needs addressee detection. Current deployments include drive-thru ordering AI where lane microphones pick up backseat passengers, car radios, and crew crosstalk; service robots and kiosks in public spaces; ambient clinical AI in exam rooms; in-cabin automotive assistants; and wearable or always-on AI in social settings. The problem is the same in each: the microphone hears more people than the device should answer.

Run SAA on your own audio before any commitment.

Evaluations are free. Tell us what you are building; we will send API keys, an eval kit, and the client SDK. Licensing is per device.

We do not share your information. See our privacy policy.