Voice pipeline layer comparison

Addressee detection vs voice activity detection

VAD hears that speech happened. Addressee detection decides whether it was meant for your device.

VAD and addressee detection ask different questions of the same audio.

VAD is cheap, fast, and belongs in every pipeline. It just cannot tell who the speech was for.

That matters in a crowded room, or on a noisy one-to-one call. Every unaddressed utterance still runs your full stack.

VAD: detects that speech is occurring. Selective Auditory Attention (SAA): detects whether that speech is directed at the device.

Layer Core question it answers When it runs Decides who is addressed?
Noise cancellation and AEC Is the audio cleaner? Pre-ASR No
Voice activity detection (VAD) Is someone speaking right now? Pre-ASR No
Speaker diarization Who spoke, and when? Typically post-ASR No
Wake word Was the trigger phrase said? Pre-ASR Trigger phrase only
SAA addressee detection Is the device being addressed, right now? Pre-ASR, in real time Yes
Noise cancellation and AEC
Core questionIs the audio cleaner?
When it runsPre-ASR
Decides addressee?No
Voice activity detection (VAD)
Core questionIs someone speaking right now?
When it runsPre-ASR
Decides addressee?No
Speaker diarization
Core questionWho spoke, and when?
When it runsTypically post-ASR
Decides addressee?No
Wake word
Core questionWas the trigger phrase said?
When it runsPre-ASR
Decides addressee?Trigger phrase only
SAA addressee detection
Core questionIs the device being addressed, right now?
When it runsPre-ASR, in real time
Decides addressee?Yes

VAD and SAA are complementary: adjacent gates, different questions.

A bystander, a second speaker, TV audio: VAD treats them all the same. Only SAA, running right after it, tells them apart.

VAD fires on every voice; SAA routes only the device-addressed one Left panel shows VAD: three speech sources (a speaker addressing the device, a bystander, and TV audio) drop vertically to a shared horizontal bus then feed a single VAD gate, producing one undifferentiated output labeled speech present yes. Right panel shows SAA: the same three sources drop to a shared bus feeding the SAA gate, which routes only the device-addressed source forward to ASR while the bystander and TV audio are dropped. The decision returns in real time, before speech-to-text, and the system fails closed when uncertain. VAD voice activity detection Speaker A Speaker B TV / radio VAD speech present: yes All three voices reach your ASR. SAA addressee detection To device Speaker B TV / radio SAA addressed → forward to ASR not addressed → drop in real time, pre-ASR, fails closed
Left: VAD fires on every voice in range. Right: SAA forwards only the device-addressed stream and holds the rest at the gate.

What VAD cannot answer becomes false activations and wasted pipeline.

VAD and SAA gate on different signals. The gap between the two is where bystander triggers and side-conversation misfires happen.

the missing signal

Addressed or not addressed

VAD fires on the person addressing the device and the bystander behind them with equal confidence.

SAA adds the one signal VAD cannot produce: whether this utterance is directed at the device.

failure mode

Suppresses on uncertainty, never passes through

When addressee intent is ambiguous, SAA suppresses rather than forwards.

VAD has no equivalent: any speech that clears it reaches your ASR and model. Every close call gets expensive.

pipeline position

VAD first, then SAA, then ASR

SAA does not replace VAD. Order matters: VAD drops noise, SAA decides addressee, ASR transcribes only what SAA forwards.

latency

Decision in real time, pre-ASR

The routing decision returns before transcription begins, well inside typical latency budgets. No GPU required.

Measured on held-out multi-party sessions.

Published in our technical paper (arXiv:2604.08412); full methodology on the research page.

Fails closed under distribution shift; cross-lingual recall is still improving.

audio and video fusion

High accuracy on the addressee decision

Adds a camera stream alongside audio for stronger context.

audio-only

Accurate enough to gate before STT

No video stream required. Hosted cloud inference.

The boundary

VAD decides that someone spoke. SAA decides who they meant it for.

The voice stack was never built to tell the two apart.

What voice AI builders ask about VAD and addressee detection.

Does VAD detect who is being addressed?

No. VAD detects that someone is speaking, nothing more. It carries no information about who the speech is meant for.

In a room with several people, VAD fires on all equally. Addressee detection alone answers who it was meant for.

Can I replace my VAD with SAA?

No. VAD gates on speech presence cheaply, dropping silence and noise.

SAA runs after VAD and decides whether that speech is addressed to the device. You need both.

Does SAA run before or after VAD in the pipeline?

SAA runs after VAD and before ASR. VAD drops silence, SAA decides addressee, ASR transcribes only what SAA forwards.

SAA does not replace VAD; it adds the decision VAD cannot make.

Why does my VAD-based pipeline misfire on background speech?

Because VAD answers whether someone is speaking, not whether the device is being spoken to.

When a bystander or background audio clears the VAD threshold, it reaches your ASR and model exactly like a real turn.

There's no signal that the speech was never meant for the device. SAA adds it before ASR runs, and fails closed when uncertain.

What does SAA return that VAD does not?

An addressee decision. SAA returns a pre-transcription routing signal: device-addressed audio is forwarded, everything else is dropped.

VAD returns only a speech-presence signal.

Request an eval or book a call.

Tell us about your voice stack. We will follow up with an eval plan for your own audio. Eval is free; licensing is per device.

Want to explore first? Try the Cloud SDK to get a feel.

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