Addressee gating for ambient clinical documentation
Ambient Scribe Addressee Gating
Your ambient documentation system captures every voice in the exam room.
Or try the Cloud SDK to get a feel.
The failure
Ambient mics capture every voice in the exam room, not just the encounter.
A far-field ambient microphone in a busy exam room captures the clinician, the patient, the patient's family, and corridor staff at once. When no layer decides which utterances belong in the record, a patient statement can appear as a clinician observation, a family remark can enter the chart as clinical history, and corridor conversation can become part of the encounter note. SAA (Selective Auditory Attention) decides which speech belongs before transcription runs.
wrong attribution
The wrong speaker enters the chart
A patient's self-reported symptom can get logged as a clinician's finding. Corrections are manual, after the fact, and often missed before the record is finalized.
bystander capture
Corridor and family voices enter the record
Ambient systems do not distinguish a bystander's comment from a diagnostic observation, so corridor speech from an adjacent bay can enter a record it has no place in. Without addressee gating, every voice in range is treated as potentially relevant.
compliance exposure
Documentation errors carry HIPAA and liability risk
A record that attributes a statement to the wrong person raises questions about the encounter. The ambient system was supposed to reduce clinician burden; a misattributed record creates avoidable HIPAA audit exposure instead.
pipeline cost
Every voice passes through your full ASR and summarization stack
Without addressee gating upstream, every voice in the room, not just the encounter, travels through your transcription and summarization pipeline, adding cost and noise at each stage.
The fix
One layer that decides which voice belongs in the clinical record.
SAA is a hosted, pre-ASR layer that decides, before speech-to-text, whether a voice is the clinician, the patient, or a bystander who should not enter the record. Ambiguous speech fails closed.
Only attributed speech routes to the ASR and ambient documentation stack you already run, with no wake word and no GPU required.
Research foundation
Built on our published technical paper.
SAA reliably identifies whether speech is addressed to the documentation system, on both audio-and-video fusion and audio-only, across held-out multi-party sessions, in the multi-talker rooms an ambient scribe operates in. Published research: arXiv:2604.08412.
Two caveats apply: (a) the system fails closed under distribution shift; (b) cross-lingual recall is a known limitation under active work.
performance
High accuracy on audio and video fusion
Evaluated across held-out multi-party sessions, the same multi-talker mix a busy exam room produces.
performance
High accuracy on audio-only
Also evaluated on audio-only, with no video stream required.
latency
Real-time decision, before STT
Runs in real time on standard hardware, no GPU. The decision returns before any word is transcribed, adding no latency to your documentation pipeline.
design principle
Fails closed by design
Ambiguous speech is not forwarded. Missing an utterance is recoverable; a misattributed clinical record may not be.
coverage
Up to 4 simultaneous speakers
Handles up to four active speakers at once, matching the clinician, patient, and family mix a busy exam room or corridor produces.
signal model
Acoustic and prosodic; no language parsing
Classifies using acoustic and prosodic signals, not transcribed words, so the decision runs entirely before ASR.
How it integrates
Where SAA sits in your ambient documentation stack.
SAA sits between your exam-room microphone and your ASR pipeline, returning an addressee decision in real time, before speech-to-text, so only addressed speech reaches the stack you already run.
No wake word. No GPU. No change to your ambient documentation vendor.
Signal in
Your exam-room audio streams to the SAA hosted service via the JavaScript/TypeScript or Python client SDK, before your ASR or scribe pipeline processes it. No model runs on your hardware. DPA terms are reviewed before any clinical audio is shared.
Addressee classification
SAA analyzes acoustic and prosodic signals, before speech-to-text, to decide whether speech belongs to the documented encounter or is bystander, corridor, or ambient audio that should not enter the record.
Route only speech addressed to the documentation system
Only utterances addressed to the documentation system pass to your ASR pipeline. Bystander speech, corridor noise, and family remarks are not forwarded; anything uncertain fails closed.
Hosted on GCP, licensed per device, DPA available
SAA runs as a hosted service on GCP and is licensed per device or deployment unit. The evaluation tier is free.
architecture
Pre-ASR, hosted, behind any ambient scribe
SAA is model-agnostic and sits upstream of your ASR and documentation platform, adding one addressee decision before transcription begins, with no change to your transcription or summarization vendor.
compliance posture
Fails closed. DPA available. No retention guarantee.
Ambiguous segments are not forwarded, and we make no retention guarantee. A DPA is available for regulated deployments; see attentionlabs.ai/dpa, or email [email protected] to begin the review before any clinical audio is shared.
Why existing tools miss it
Speaker diarization and VAD do not gate by addressee before transcription.
Each layer below solves a real problem in the ambient audio pipeline, but none decides before ASR whether an utterance belongs in the clinical record. That gate has to run upstream, or every voice in the room reaches the part of the pipeline that acts on it.
| Layer | What it answers | When it runs | Gates by addressee before ASR? |
|---|---|---|---|
| Noise cancellation and AEC | Is the audio cleaner? | Pre-ASR | No |
| Voice activity detection (VAD) | Is someone speaking? | Pre-ASR | No |
| Beamforming / spatial filtering | Where is the sound coming from? | Pre-ASR | No |
| Speaker diarization | Who spoke, and when? | Post-ASR | No |
| Speaker identification | Is this a known voice? | Post-ASR | No |
| SAA addressee gating | Is this utterance addressed to the documentation system, and should it be transcribed? | Pre-ASR, in real time | Yes |
Clinical AI teams and ambient scribe platforms where every word in the record must be attributed to the right person.
Common questions
What clinical teams ask before deploying.
How does wrong-speaker attribution happen in an ambient scribe?
An ambient scribe captures all room audio without deciding which utterances belong in the chart. A patient statement can be logged as a clinician's observation, a family member's remark can enter the chart as a diagnostic observation, or corridor talk can enter the encounter note, because no gate runs before transcription.
Does audio leave the facility, and how is protected health information handled?
Audio streams to attention labs hosted infrastructure on GCP to produce the addressee decision, which returns in real time, before speech-to-text. A Data Processing Agreement is available for regulated deployments, and we make no retention guarantee. Email [email protected] to begin that review before any regulated audio reaches the hosted layer.
Is a Data Processing Agreement available for regulated deployments?
Yes, for health system and ambient documentation platform deployments. We treat the DPA discussion as part of evaluation, not a post-signature step. Request an eval or book a call and we will route you to compliance review first.
How does SAA decide which speech belongs in the clinical record?
SAA analyzes acoustic and prosodic signals, not words, to decide whether speech belongs to the documented clinician-patient encounter or is bystander, corridor, or ambient audio, before ASR runs. It does not rely on a wake word. Segments classified as bystander, corridor, or ambient are not forwarded to the transcription pipeline. It fails closed under distribution shift, and cross-lingual recall is a known limitation under active work.
What happens when SAA is uncertain whether speech is addressed to the documentation system?
SAA fails closed: the audio segment is not forwarded to the transcription pipeline. In a regulated clinical context, missing an uncertain utterance is recoverable; forwarding a bystander statement into the record may not be. That tradeoff is the conservative default by design.
How do we evaluate SAA on our own consented room audio?
The evaluation tier is free. Request an eval and we will assess audio from your own environment, with DPA terms reviewed before any regulated audio is shared. You also get API keys, an eval kit, and the client SDK to measure addressee decisions yourself.
Get in touch
Book a call or request an eval.
Tell us about your ambient documentation workflow, the clinical environment, and where speaker attribution is breaking down.
We will follow up with next steps for your evaluation, including a DPA review before any regulated audio is shared.
Or try the Cloud SDK to get a feel.
Last updated June 26, 2026