Free Selective Auditory Attention (SAA) Cloud Eval: test addressee detection on your own audio.

SAA scores addressed vs. not-addressed speech on your data, not a benchmark.

Everything you need to run the eval on your own audio.

Credentials land within 24 hours for qualified projects. No call required to start.

Stream your own audio through the hosted SAA API and see what it scores as addressed vs. not.

hosted API key

Cloud API access to server.attentionlabs.ai

Key-gated access to the SAA inference endpoint, elevated rate limits, no model on your hardware.

eval kit

Sample audio and deterministic scoring scripts

Curated multi-party audio plus scripts that score false-trigger rate and addressee accuracy on your own audio.

client SDK

JavaScript/TypeScript and Python SDKs

Thin SDKs that stream audio from your pipeline to the hosted endpoint, the same artifact used in production.

timeline

48 to 72 hours to a scored result

If you hit a blocker, we respond the same business day. A DPA is available on request for regulated environments.

From your audio to a scored result in three steps.

Score your own audio on the same hosted, real-time, pre-ASR path as production.

The SAA Cloud Eval pipeline: from your audio to per-segment scores Your environment audio streams to the SAA Cloud API. Per-segment addressee-detection decisions are returned in real time before ASR runs. The output is a scored result on your own data: which segments are addressed to the device and which are not, plus false-trigger rate and addressee accuracy on your environment. Your environment audio lane, lobby, call, cabin, or any multi-party environment streams to API SAA Cloud API addressee inference pre-ASR · real-time per segment · fails closed scores returned Per-segment addressee-detection scores addressed / not addressed · on your own audio
The green arrow marks addressed audio forwarded to your ASR. Everything the API classifies as not addressed or uncertain is marked for suppression.
addressed seg 0:04 to 0:07, conf 0.94
not addressed seg 0:09 to 0:11, conf 0.88
uncertain / closed seg 0:13 to 0:14, conf 0.51

Simulated output. Actual scores are computed on your own audio.

What the published research says, and what it does not.

SAA scores addressed vs. not-addressed speech across held-out tests; audio-plus-video beats audio alone (arXiv:2604.08412).

Cross-lingual recall is a known limitation under active work. Published results are a starting point, not your number.

latency

Pre-ASR, in real time

The eval measures this on your integration, verified against your pipeline budget, not a lab number.

design principle

Fails closed, not open

Ambiguous speech is held, not forwarded; the default under uncertainty is silence. The eval logs every closed abstention.

coverage

Real multi-speaker environments, no wake word

Runs on multi-party audio: lane, cabin, call bridge, lobby. Acoustic-level and language-agnostic.

Tell us about your environment. We will send credentials.

Submit the form and qualified projects receive credentials within 24 hours. Note DPA requests in the use-case field.

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

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

Three steps from form to scored result.

The path is short on purpose. The scored result is the artifact that makes the follow-up conversation useful.

step 01

We review and provision within 24 hours

We confirm we can support your integration, then provision credentials and documentation.

If you need a DPA first, we process both together. If a call is needed before credentials, we will say so.

step 02

You run the eval on your own audio

Stream audio through the SAA API and run the provided scoring scripts.

A meaningful result means real addressed and not-addressed segments scored on your own audio, not a demo.

step 03

We discuss the result and scope next steps

A member of our team reviews the numbers with you and scopes production if the result is positive.

Production requires a separate per-device commercial license.

Built for

Engineering and product teams who need a real number on their own audio before they decide. Not a benchmark. Not a demo reel. Your data, your environment, your result.

What teams ask before requesting the eval.

Is the SAA Cloud Eval really free?

Yes. Credentials, sample audio, scoring scripts, and the SDK are free. Production needs a per-device license.

What does the eval include?

A cloud API key, sample audio and scoring scripts, the client SDK, and setup docs to guide you.

How long does the evaluation take?

Most teams reach a scored result within 48 to 72 hours of receiving credentials.

What happens to the audio I send through the API?

Audio streams to the hosted endpoint; decisions return before ASR, and audio is not retained after inference.

Is a data processing agreement available?

Yes, for regulated deployments. Request it in the eval form or email [email protected].

What is the difference between the eval and a production license?

The eval is evaluation-only, cloud-hosted and key-gated. Production needs a separate per-device license.

What happens after I submit the form?

Qualified projects get credentials within 24 hours. We follow up the same business day if you are a fit.

Request the eval or book a call.

The eval is the fastest path to a number on your own data. Prefer to talk first? Book a call.