Addressee detection for voice AI builders

Voice separation finds every voice, not who it's for.

Selective Auditory Attention (SAA) adds that missing decision, in real time, before ASR runs, and fails closed when unsure.

What is the cocktail party problem?

The cocktail party problem describes how people follow one voice among many in a noisy room. E. Colin Cherry named it in 1953.

Voice separation only clears the mix. A bystander saying your wake phrase to a friend still produces a clean signal, and ASR will act on it.

SAA answers what separation cannot: which voice is addressed to the device. It runs before ASR, in real time, and fails closed when unsure.

In the original 1953 formulation, the listener is always the intended recipient. In a deployed voice AI product that assumption breaks constantly: a drive-through assistant hears the driver, the back seat, the crew, and the radio at once; a lobby robot is surrounded by people talking to each other, not to it; an in-cabin assistant receives speech from every seat. The question is not only which voice, but whether any voice is addressing the device at all.

That second question, directed at the device or not, is the engagement-control problem. It needs a dedicated layer between the microphone and the recognition pipeline. Source separation, voice activity detection, and wake words each solve a slice of it. None of them close the loop on addressee intent, in real time, before transcription begins.

No layer in your voice stack knows who is being addressed.

Separation, VAD, diarization, and wake words each solve a different piece of the audio problem.

Layer What it answers When it runs Decides who is addressed?
Source separation Which individual voices are in the mix? Pre-ASR No
Voice activity detection (VAD) Is someone speaking right now? Pre-ASR No
Speaker diarization Who spoke, after the fact? Post-ASR No
Wake word detection Was the trigger phrase said? Pre-ASR Trigger phrase only
SAA addressee detection Is the device being addressed right now? Pre-ASR, real time Yes

Where SAA sits: after separation, before ASR.

It decides, per voice, whether that voice is addressed to the device or to someone else in the room.

Two-stage diagram: voice separation then SAA addressee detection Four colored waveforms representing competing voices converge on a separation stage that untangles them into four clean lanes. A second highlighted stage, the SAA gate, selects only the addressed lane to forward to ASR and marks the other three as not addressed. STAGE 1 STAGE 2 Separation (prior art) untangles voices into clean lanes SAA which lane is addressed? real time · pre-ASR fails closed addressed → ASR not addressed not addressed not addressed
Four competing voices enter the separation stage, which produces clean lanes. The SAA gate selects the lane addressed to the device and forwards it to ASR. The other three are dropped. When SAA is uncertain, it fails closed and forwards nothing.
01

Drop in before your ASR, no pipeline changes

Audio and optional video stream to the hosted SAA service via a thin SDK (JavaScript/TypeScript or Python).

No model runs on your hardware. No GPU. No wake word. SAA inserts as a single stage between your microphone and ASR.

02

Addressee decision in real time, from acoustic signal alone

SAA analyzes acoustic and prosodic signals in the raw audio frames and returns addressed or not addressed, before ASR runs.

No trigger phrase. No text. No post-hoc LLM call.

03

Only addressed audio reaches your model; everything else is suppressed

Addressed audio forwards to your ASR and language model. Non-addressed speech is suppressed.

Uncertain audio fails closed: nothing is forwarded on a guess. You keep your existing ASR, model, and pipeline.

Reliable on the addressee decision, audio-only or fused with video.

Measured across held-out multi-party sessions with multiple concurrent speakers. Full methodology: arXiv:2604.08412.

Two caveats: it fails closed under distribution shift, and cross-lingual recall is a known limitation under active work.

performance

High accuracy with audio and video

Combining audio and video gives SAA the clearest signal for the decision.

performance

Accurate on audio alone

No camera required. The same decision runs on audio alone.

latency

Real-time decision

The routing decision returns in real time, before ASR begins processing.

design principle

Fails closed by design

Ambiguous speech is suppressed, not passed through. Silence beats a forwarded guess.

Every shared acoustic space is a cocktail party.

Different environments, same failure: a voice agent in a real acoustic space can hear audio it was never meant to act on, in one-to-one calls too.

STT, LLM, and TTS run on every second received, including audio addressed to no one. SAA gates it before ASR, so that compute never runs.

01

Drive-through ordering

Driver, backseat, crew, radio: four clean channels, and none of them are labelled "for the AI."

Without an addressee gate, the agent acts on whichever voice is loudest.

02

Service robots

A robot in a lobby hears dozens of conversations it is not part of. A wake word fires on bystanders; addressee detection does not.

03

In-cabin voice assistants

Beamforming and seat routing separate audio spatially, but not by intent.

When two seats speak at once, neither one says which occupant the device should answer.

04

Ambient clinical AI

Clinician, patient, family members: without addressee detection, the scribe captures all of them and attributes utterances to the wrong role.

Wrong attribution here is a documentation error in the chart.

The bottom line

Your agent does not have a separation problem.

What voice AI builders ask about the cocktail party problem.

What is the cocktail party problem?

The cocktail party problem describes how people follow one voice among many in a noisy room. E. Colin Cherry named it in 1953.

For voice AI it has a second layer: the device must decide which voice, if any, is addressed to it, not just which voices are present.

Does source separation solve the cocktail party problem for voice AI?

Separation isolates voices. A cleanly isolated voice can still be addressed to another person, not the device.

SAA adds the missing decision: in real time, before ASR runs, it determines whether each voice is device-directed.

How is addressee detection different from voice separation?

Voice separation answers which voice is present. Addressee detection answers whether that voice is talking to the device.

Separation untangles mixed audio into clean channels. SAA classifies each channel by intent, using acoustic and prosodic signals rather than content, and forwards only the addressed one to ASR.

No text. No trigger phrase.

Can SAA handle four people talking at once?

Yes. SAA is designed for real multi-party environments with multiple concurrent speakers.

It decides which, if any, of the active voices is addressed to the device, and forwards only that audio onward.

When it cannot determine with confidence that any utterance is device-addressed, it fails closed: audio is not forwarded rather than passed through on a guess.

Where does this matter in production?

Any voice AI device in a shared acoustic space faces this: drive-throughs, lobbies, vehicle cabins, and exam rooms all carry voices the device was never meant to act on.

In each case the fix is a pre-ASR addressee gate that routes only addressed audio to the model.

Run SAA on audio from your own deployment.

Describe your product and the acoustic environment it runs in. We will follow up with next steps for a free evaluation; licensing is per device.

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

Or try the Cloud SDK to get a feel.