Voice pipeline layer comparison

Addressee detection vs noise suppression

Noise suppression cleans the audio. Addressee detection decides whether it was meant for your device.

No, addressee detection is not a noise filter. It classifies intent.

Selective Auditory Attention (SAA) classifies who speech was addressed to, then forwards it to ASR, or holds it back.

Noise suppression: cleans the audio. SAA: decides who it was meant for.

The bystander's voice is clear. That is the problem.

At a drive-thru lane, a customer orders while a crew member calls out to a coworker nearby. Both voices arrive clean.

Noise suppression and beamforming can't tell who a clear voice was meant for. Only intent classification can.

Signal processing cleans audio. Addressee detection decides where it goes.

Each layer below answers a different question. The output type shows why.

Layer Core question it answers Output type Decides who is addressed?
Noise suppression and AEC Is the audio cleaner? Transformed audio No
Beamforming Is the target direction prioritized? Spatially filtered audio No
Voice activity detection (VAD) Is someone speaking right now? Binary speech-present signal No
SAA addressee detection Is the device being addressed, right now? Routing decision Yes

Signal processing reads the waveform. Addressee detection reads intent.

SAA models prosody, gaze cues, and multi-speaker spatial patterns. These are inputs the signal-processing stack never sees.

complementary, not competing

They belong in the same pipeline, in order

SAA runs after noise suppression and VAD, right before ASR, adding the one decision neither stage makes.

failure mode without SAA

Every unaddressed voice reaches your ASR

Without SAA, any voice that clears VAD triggers a full turn, even when no one addressed the device.

This includes one-to-one setups with background voices. SAA gates before ASR, so that cost is never paid.

Environments where the audio is clear and the intent is not.

drive-thru ordering

Customer vs crew cross-talk

Customer and crew voices often arrive at identical volume and clarity.

SAA classifies intent and holds the crew voice before it reaches the ordering model.

car cabin

Driver vs passenger commands

In a quiet cabin, any passenger utterance arrives clearly.

SAA classifies who the speech was addressed to, so the assistant responds only when it is.

contact center

Agent side-conversations

When an agent's mouthpiece cover slips, the coworker's voice arrives cleanly.

There is no noise to suppress.

SAA classifies it as not device-directed and holds it before transcription or sentiment analysis runs.

Measured on real multi-party sessions where signal quality is not the bottleneck.

Published at arXiv:2604.08412 and evaluated on multi-party sessions where intent, not audio, was the variable.

Two caveats: the system fails closed under distribution shift, and cross-lingual recall remains a known limitation.

audio and video fusion

Highest accuracy with audio and video

Adding a camera feed further improves reliability when one is available.

audio-only

Reliable on audio alone

Works on audio alone, no camera required, via hosted cloud inference.

The distinction

Clean audio and correct addressee are two different questions.

The voice stack was never built to answer who the speech was for.

What voice AI builders ask about noise suppression and addressee detection.

Does noise suppression decide who is talking to the device?

No. Noise suppression cleans the audio: less noise, less echo, better signal quality.

It never decides who the speech was meant for. That is what addressee detection does.

Can beamforming replace addressee detection?

No. Beamforming prioritizes audio from a spatial direction.

It can suppress voices from behind the mic array, not two from the same direction.

In a drive-thru, car, or meeting room, filter and keep often stand in the same spot.

Spatial filtering cannot resolve communicative intent.

Where does SAA sit in the pipeline relative to noise suppression?

SAA runs after noise suppression and VAD, but before ASR.

Order: noise suppression, VAD (drops silence), SAA (decides addressee), then ASR.

SAA adds the one decision the other stages don't make: who the speech is for.

Does SAA require clean audio to work?

SAA is built for real-world environments where audio quality varies.

We recommend upstream noise suppression for loud deployment environments.

When voices are already clear, addressee is the harder problem, and SAA solves it standalone.

Request an eval or book a call.

Tell us about your voice stack for a free evaluation on your own audio. Licensing is per device.

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