Voice pipeline fundamentals
What is pre-ASR gating?
Deciding what reaches your ASR, before ASR runs.
Pipeline position
Pre-ASR gating is the decision VAD never made.
A voice AI pipeline runs in stages: noise cancellation, voice activity detection (VAD), ASR, then the language model.
VAD forwards everything it detects. It never asks who the speech was meant for.
Every utterance that clears VAD burns ASR, model, and TTS cost, addressed or not. A pre-ASR gate drops the unaddressed share before transcription starts.
VAD: detects that speech is present, and forwards it. Pre-ASR gate (Selective Auditory Attention, SAA): decides whether that speech is addressed to the device before forwarding it to ASR.
| Pipeline stage | Without pre-ASR gating | With pre-ASR gating (SAA) | Where the cost lands |
|---|---|---|---|
| VAD | Fires on all detected speech | Fires on all detected speech | Same either way |
| Addressee decision | None. All speech continues to ASR. | SAA returns forward or suppress in real time, before ASR | Gate fires here |
| ASR call | Made for every utterance VAD passes | Made for addressed speech only | ASR cost removed |
| LLM / model call | Made for every ASR transcript | Made for addressed transcripts only | Model cost removed |
| False trigger | Proportional to bystander speech in range | Suppressed pre-ASR, before compute runs | Prevented at source |
Cost and resilience
Every utterance costs money, addressed or not.
ASR, model, and TTS calls are metered per use. The bill does not know which share of that audio was never meant for the device.
This applies to a quiet one-to-one conversation as much as a crowded room. Filtering before transcription is where it is cheapest to stop.
Where it applies
Background speech counts too, not just crowded rooms.
drive-through lanes
Order-taking in multi-talker lanes
A drive-through lane hears order-takers, customers, and passing conversations at once. Only speech directed at the terminal should reach the model.
in-vehicle and cabin AI
Car cabins with driver and passengers
A vehicle assistant hears driver and passengers at roughly equal volume. Cabin chatter clears VAD the same way a real command does, so gating is what keeps it silent.
meeting rooms and conferencing AI
Rooms where several people are speaking
Meeting room AI captures every participant, including side conversations. Gating lets it respond only when directly addressed, no wake word needed.
contact centers and agent assist
Agent stations with ambient colleague speech
Agent stations pick up nearby colleagues as well as the agent. A pre-ASR addressee gate routes only agent-addressed audio to the assist model.
How SAA works
SAA gates on addressee detection, not a keyword or energy threshold.
SAA evaluates each audio segment in real time and returns forward or suppress before ASR runs. Trained on real multi-party audio, it suppresses when uncertain.
audio and video fusion
High accuracy on the addressee decision
Evaluated on held-out multi-party sessions combining audio and video. Published at arXiv:2604.08412. Cross-lingual recall is a known limitation under active work.
audio-only
Accurate from audio alone
Real multi-speaker audio, no video stream required. Works over telephony and audio-only pipelines with hosted cloud inference.
no wake word required
Natural speech, no trigger phrase
SAA gates on addressee intent inferred from audio and contextual signals. Speakers address the device naturally.
works behind any model
ASR and LLM agnostic
SAA forwards audio, not transcripts, sitting between the mic and your ASR of choice. Swap ASR or LLM with no change to the gating layer.
VAD answers whether speech is present. The addressee decision is what the existing voice stack was not built to make.
Common questions
What voice AI engineers ask about pre-ASR gating.
What is pre-ASR gating?
Pre-ASR gating decides, before ASR runs, whether to forward or suppress an audio segment based on who it is addressed to.
How does pre-ASR gating differ from voice activity detection?
VAD detects that someone is speaking; it cannot tell who they are speaking to. Pre-ASR gating answers that second question: is this speech addressed to the device.
Does pre-ASR gating require a wake word?
No. A wake word is one form of gate, but needs a trigger phrase. SAA reads natural speech and decides addressee without one.
Which environments benefit most from pre-ASR gating?
Any space with ambient speech qualifies: drive-through lanes, car cabins, meeting rooms, contact centers. VAD alone is enough only in a quiet single-speaker setting.
What does SAA return as its pre-ASR gating decision?
SAA returns forward or suppress for each segment, in real time before transcription runs. When uncertain, it suppresses rather than forwards.
Get in touch
Request an eval or book a call.
Tell us about your voice stack. We will follow up with next steps for a free evaluation. Eval is free; licensing is per device.
Last updated June 29, 2026