Pre-ASR engagement control for voice AI

Addressee detection vs wake words

Addressee detection decides who a device is being spoken to from natural speech, no trigger required, before ASR runs.

A workaround vs. the actual decision.

A wake word reduces idle listening. Selective Auditory Attention (SAA) decides whether the device was actually addressed, trigger or not.

wake word detection

The workaround: a user-performed protocol

Any audio with the trigger phrase unlocks the pipeline, broadcast or not. It confirms the phrase, not who it was for.

A fixed phoneme match cannot tell a side conversation from a real command, so a phrase quoted on TV or spoken in another room passes the same as an intended trigger.

addressee detection (SAA)

The decision: is this utterance for the device?

SAA classifies addressee from acoustic signals alone, before ASR, no trigger needed. Fails closed when uncertain.

No GPU is required. Non-addressed audio is never forwarded downstream, and a DPA is available for regulated deployments.

What each layer decides, and whether it decides addressee.

VAD, wake word, and diarization each answer a different question; none decide addressee.

Each row is a real layer in the pre-ASR or post-ASR pipeline, including whether a trigger phrase is required. Only one layer decides addressee continuously.

Layer What it decides When it runs Trigger phrase required? Decides addressee?
Voice activity detection (VAD) Is someone speaking? Pre-ASR No No
Wake word Was the trigger phrase spoken? Pre-ASR Yes Phrase-gated only
Speaker diarization Who spoke and when? Post-ASR No No
SAA addressee detection Is the device being addressed right now? Pre-ASR, in real time No Yes

How each approach handles the same stream of speech.

The timeline below shows the same audio stream handled by two models. In the wake-word model, every interaction requires a trigger phrase before the command, and background audio containing the phrase causes a false unlock. In the addressee-detection model, natural speech is classified continuously: addressed turns are forwarded without a trigger, non-addressed turns are dropped, and uncertain turns fail closed.

Timeline comparison: wake word model vs SAA addressee detection Top track: wake word model. A trigger phrase is required before each command. A background phrase causes a false unlock. A natural phrase with no trigger is dropped. Bottom track: SAA addressee detection. No trigger required. Addressed speech is forwarded. Non-addressed speech is dropped. Uncertain speech fails closed. Wake word model Trigger phrase required before every command T Forwarded trigger + cmd ! False trigger background phrase × Missed no trigger phrase T Forwarded trigger + cmd Addressee detection (SAA) No trigger required. Fails closed under uncertainty. A Forwarded addressed S Dropped side conv. A Forwarded addressed ? Fails closed uncertain A Forwarded addressed Addressed, forwarded Not addressed, dropped Uncertain, fails closed
Wake words miss natural speech and false-trigger on background audio. SAA classifies directly and fails closed when uncertain.

Four things a wake word cannot do.

A wake word detects a phoneme sequence. SAA decides conversational intent. The two are not in competition: they can run together, with the wake word as an optional explicit-activation layer and SAA providing continuous addressee gating across the session.

only SAA decides

Natural speech, no trigger required

Natural speech with no keyword is still forwarded. A wake word only recognizes the trigger itself.

only SAA decides

Trigger phrase in background speech

SAA suppresses the trigger phrase in audio not addressed to the device, a distinction wake words cannot make.

only SAA decides

Continuous addressee state across speakers

Every speaker gets an independent addressee decision each turn, not only right after a trigger. In a multi-party room, SAA maintains this signal for the full session.

only SAA decides

Ambiguous utterances, fail closed

When acoustic evidence is ambiguous, SAA does not guess. Audio is not forwarded; ASR never receives it.

Measured performance on real multi-party audio.

SAA reliably classifies addressee from audio-video fusion or audio alone, on held-out multi-party sessions (arXiv:2604.08412).

It fails closed under distribution shift; cross-lingual recall is a known limitation under active work.

Every audio segment sent to ASR, LLM, and TTS costs compute, addressed or not. SAA drops it before the pipeline runs.

In QSR, cross-talk false-triggers open-mic incidents; SAA suppresses them pre-ASR. In service robotics, SAA replaces an often-impractical trigger phrase.

performance

Reliable on audio alone or audio-video fusion

Both modalities are evaluated on the same benchmark. Audio alone needs no video stream; fusion adds spatial context when a camera is present.

latency

Real-time decision, pre-ASR

No GPU required. Runs under any ASR or model stack, on any device.

design principle

Fails closed by design

Ambiguous speech is suppressed, not passed through. Under uncertainty the default is silence, not a false activation, a behavior wake words have no equivalent for.

What teams ask before switching architectures.

Can I drop the wake word entirely?

It depends on your product design. SAA decides addressee on any natural utterance, no trigger needed. Some teams drop the wake word entirely; others run a hybrid architecture where SAA provides continuous addressee gating and the wake word stays as an optional explicit-activation layer. Nothing about SAA requires removing an existing wake word.

Why do wake words misfire?

A wake word fires whenever its trigger phonemes appear, broadcast or genuinely addressed. A phrase quoted on TV, spoken in a side conversation, or repeated by another person can all trigger a false activation. SAA adds the addressee judgment phoneme matching lacks, so it suppresses activations where the trigger phrase was present but the speech was not directed at the device.

Does SAA work alongside an existing wake word?

Yes, with no changes to your stack. SAA runs upstream, downstream, or parallel to any wake-word detector, gating everything the trigger does not. In a hybrid deployment the wake word provides explicit user activation while SAA suppresses ambient speech that contains the trigger and handles natural-speech turns after activation.

Is addressee detection slower than a wake word?

No. SAA decides before ASR begins, the same point in the pipeline as a wake word, but from a richer acoustic signal rather than a narrow phoneme match. The combined latency of SAA plus your ASR and model stack is comparable to a wake-word-gated pipeline.

What about privacy if there is no wake word?

SAA streams audio to a hosted inference endpoint; non-addressed audio is never forwarded downstream and is not retained after inference. A DPA is available for regulated deployments.

How addressee detection relates to other audio pipeline layers.

Each comparison page explains one distinct architectural decision and how SAA relates to it. Together they define the addressee-detection category.

A path to natural, trigger-free interaction.

Tell us about your stack and we will follow up with a free evaluation on your own audio.

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

Or try the Cloud SDK to get a feel, or email us directly: [email protected]