Pre-ASR addressee gating for contact-center voice AI

Your agent hears the whole call.

A pre-ASR addressee gate cuts false escalations and cleans the signal for containment rate and handle time.

Your agent has no signal for who is addressing it.

A supervisor aside, a household voice, hold music: none of it is the customer. Each one can still trip an escalation.

At scale, even a low false-escalation rate degrades containment and inflates handle time.

The recognizer is not the problem. The missing piece is a decision, before ASR runs, about who a turn is addressed to.

A contact-center call bridge with three audio sources reaching the agent runtime The customer voice, a supervisor whisper, and background side conversation all reach the call bridge, each on its own path. With no addressee gate, the agent runtime acts on all three identically, producing false escalations and false barge-in that inflate handle time and degrade containment rates. Customer voice principal speaker Supervisor whisper coaching aside Side conversation background speech Agent runtime · no addressee gate Acts on all three turns false escalation · false barge-in handle-time inflation
Lime marks the real customer. Without an addressee gate, all three paths reach the agent runtime the same way.

false escalation

The supervisor aside becomes a customer command

A coaching whisper on a three-way bridge fires an escalation trigger. The call routes to a live agent.

In healthcare and financial lines, the misrouting creates compliance risk, not just noise.

false barge-in and missed interruption

Background speech stops the agent. Real interruptions get missed.

Hold music, a household voice, or a side conversation trips the VAD. The agent cancels its response mid-turn, pauses, then awkwardly resumes.

The real interruption that mattered gets buried in the same noise and missed. Handle time inflates, and the root cause is no addressee gate before ASR.

One hosted layer, before ASR, that decides who a turn is addressed to.

Supervisor asides, hold music, household voices, and side conversations are suppressed. Only turns addressed to the agent reach your existing runtime.

When Selective Auditory Attention (SAA) is uncertain, it fails closed and forwards nothing. No wake word, no GPU, no change to your ASR or LLM.

Published research on multi-party call audio.

Not just claimed: measured on held-out multi-party audio and published for scrutiny.

addressee precision

Reliable on the addressee decision, audio-only or with video

Measured on held-out multi-party sessions comparable to live call audio. Published at arXiv:2604.08412.

Fails closed under distribution shift.

latency

In real time, before ASR runs

The addressee decision returns before the recognizer processes any audio.

Fits inside a live call latency budget with no perceptible delay.

failure mode

Fails closed, not open

When SAA is uncertain, the default is silence. The turn is not forwarded to the agent runtime.

No false escalation can originate from a suppressed turn.

coverage

Multi-party call audio, several speakers at once

Works across three-way bridges and caller environments with background voices, with several speakers active at once.

signal model

Acoustic and prosodic signals, no language parsing

No wake word. Language-agnostic at the acoustic level.

Cross-lingual recall parity is a known open item, under active work.

deployment

Hosted cloud, DPA available, per-device licensing

SAA runs as a hosted service. Call audio streams to the inference endpoint; an addressee decision returns.

A Data Processing Agreement is available for regulated lines including healthcare and financial services.

Licensed per device, not per call.

One SDK call between your call transport and ASR. Nothing else changes.

Thin client SDKs in JavaScript/TypeScript and Python. LiveKit, Pipecat, and Twilio transports supported.

How SAA routes only addressed call turns to the agent runtime Raw call audio containing all voices on the channel streams in. SAA decides in real time, before ASR, whether a voice turn is addressed to the agent. Addressed turns are routed to ASR and the agent runtime. Turns that are not addressed, or uncertain, are not forwarded; the system fails closed. Call audio all voices on the channel SAA addressee decision real time, pre-ASR Addressed turn → your ASR and agent runtime Not addressed fails closed · not forwarded
SAA sits before ASR. Addressed turns reach the agent runtime; everything else stops at the gate.
01

Audio in from your existing transport

Audio from your phone transport or CPaaS bridge streams to SAA before ASR runs.

No change to your transport, microphone configuration, or telephony stack.

02

Addressee decision per voice turn

SAA analyzes acoustic and prosodic signals for each voice turn. No wake word required. No GPU required.

In real time it decides whether the turn is addressed to the agent, or to someone else on the bridge or in the caller's room.

03

Route only addressed turns to your agent

Only turns addressed to the agent reach your ASR and agent runtime. Supervisor asides, hold music, household voices, and side conversations are suppressed.

Under uncertainty SAA fails closed and forwards nothing.

04

Real-time on a live call, at scale

The agent's response rhythm stays the same. It just stops acting on the wrong speaker.

architecture

Pre-ASR, hosted, model-agnostic

Client SDKs stream audio to server.attentionlabs.ai and get an addressee decision before transcription starts. SAA sits in front of any ASR and any agent runtime you already run.

access

Free eval, one call type, one week.

Run SAA against audio from your own environment, or try the Cloud SDK first to get a feel.

Request an eval

VAD and diarization tell you someone spoke. Not whether it was meant for the agent.

Four layers already sit in your pipeline. Only one answers the addressee question.

Layer What it answers When it runs Answers "is this turn addressed to the agent?"
Energy-threshold VAD + endpointing Is someone speaking? Has the turn ended? Pre-ASR No
Noise suppression and AEC Is the audio cleaner? Pre-ASR No
Speaker diarization Who spoke and when, after the fact? Post-ASR No
Voice biometrics / caller authentication Is this a known enrolled voice? Post-ASR No
SAA addressee detection Is this turn addressed to the agent, right now? Pre-ASR, in real time Yes
Layer
Energy-threshold VAD + endpointing
What it answers
Is someone speaking? Has the turn ended?
When it runs
Pre-ASR
Answers "is this turn addressed to the agent?"
No
Layer
Noise suppression and AEC
What it answers
Is the audio cleaner?
When it runs
Pre-ASR
Answers "is this turn addressed to the agent?"
No
Layer
Speaker diarization
What it answers
Who spoke and when, after the fact?
When it runs
Post-ASR
Answers "is this turn addressed to the agent?"
No
Layer
Voice biometrics / caller authentication
What it answers
Is this a known enrolled voice?
When it runs
Post-ASR
Answers "is this turn addressed to the agent?"
No
Layer
SAA addressee detection
What it answers
Is this turn addressed to the agent, right now?
When it runs
Pre-ASR, in real time
Answers "is this turn addressed to the agent?"
Yes
Built for

CCaaS and CX teams whose containment rate depends on answering the right person.

What CCaaS and CX platform teams ask before integrating.

Why do contact-center voice agents falsely escalate?

When a supervisor coaches the customer mid-call, or ambient room noise reaches the channel, the agent runtime has no signal that the turn was not addressed to it.

Energy-threshold VAD fires on any sound above a noise floor, and the escalation trigger fires on content it was never meant to receive.

The root cause is not the escalation classifier: it is a missing decision about who a turn is addressed to. SAA adds that decision before ASR runs.

How is SAA different from our VAD plus endpointing model?

VAD answers whether someone is speaking. SAA answers whether that speech is addressed to the agent.

VAD fires on hold music, background conversation, and a supervisor's coaching aside. None of that carries an addressee signal.

SAA is the classification layer above your VAD. It is a single SDK integration that does not change your ASR or LLM.

How does SAA handle supervisor barge-in and 3-way calls?

On a three-way bridge, SAA classifies each voice turn by addressee in real time, before ASR runs.

A supervisor's whisper is tagged as not addressed and withheld. A turn directed at the agent is forwarded.

When SAA is uncertain, it fails closed and withholds the audio. The agent runtime receives only turns addressed to it.

Is audio sent to the cloud, and is there a DPA for regulated lines?

SAA is a hosted cloud service. Call audio streams through the client SDK to server.attentionlabs.ai and an addressee decision returns in real time.

SAA fails closed: unaddressed and uncertain audio is never passed onward. A Data Processing Agreement is available for regulated deployments.

How fast is the evaluation and what does it measure?

The evaluation tier is free. A typical single-call-type evaluation runs in one week and includes API keys, an eval kit, and the client SDK.

You run SAA against audio from your own call types and receive per-turn addressee decisions.

The evaluation measures false-escalation reduction, per-turn addressee precision, and decision latency, all on your own audio.

How is SAA licensed for high call volume?

SAA is licensed per device, not per call and not metered by call volume.

Contact-center platform teams typically start with one call type and expand across the fleet after a successful evaluation.

Request an eval to run SAA on your own call audio, or book a call to discuss commercial terms.

Book a call or run an eval on your own call audio.

Tell us about your call types and what you are trying to fix: false escalation rate, handle time, or compliance on a regulated line.

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

Related: how to reduce false triggers in voice AI · what background speech costs a voice agent