Addressee Detection for Aviation Voice AI

Gate kiosk and check-in microphones capture every sound in range: the boarding PA calling groups overhead, a companion standing next to the transacting passenger, crowd chatter at 80 to 90 decibels, and adjacent kiosk sessions running concurrently. Without a layer that decides who is addressing the device, all of it reaches the ASR as potential user input. The PA announcement is not a recognizer failure; the system simply has no signal for which speaker was addressing the kiosk. SAA adds that decision before transcription runs.

The recommended entry environment is gate or check-in. Both are bounded installations with measurable interaction volumes, no cabin certification burden, and a business unit (Operations or CX) that owns a clear KPI such as false trigger rate or passenger handle time. The cabin seatback environment is a valid second step after a gate or check-in evaluation closes successfully.

Weeks 1 and 2 cover integration and baseline. You provide representative audio from the target environment, typically 10 to 20 hours of interactions including PA announcements, ambient crowd, and true addressed sessions. We deliver API keys, a client SDK in Python or C++, and the eval kit. Integration is a single pre-ASR stage in your existing pipeline. Weeks 3 and 4 run the SAA-gated pipeline against the same audio, measuring four agreed metrics: false trigger rate, addressed-speech F1, decision latency, and PA noise tolerance. Weeks 5 and 6 produce the written go/no-go report with per-metric results versus baseline, distribution shift observations, and a recommendation for the next environment. The evaluation fee is convertible into pilot credit or a strategic option.

Yes. SAA is a hosted cloud service. Audio streams to the inference endpoint at server.attentionlabs.ai and the addressee decision returns in real time, before speech-to-text. SAA fails closed when confidence is below threshold, so unaddressed audio and uncertain audio are not forwarded downstream. A Data Processing Agreement is available for regulated environments and is standard for aviation deployments. Request one at the time of evaluation.

The evaluation measures four metrics against your existing pipeline baseline: false trigger rate, the fraction of non-addressed audio events such as PA, ambient crowd, bystander, and companion that reach ASR; addressed-speech F1, the accuracy on audio genuinely directed at the device; decision latency, the time from audio frame ingress to routing verdict, which must return in real time, before speech-to-text; and PA noise tolerance, a specific subset test with concurrent PA announcements running at measured decibel levels representative of your environment. Low-confidence abstentions, the events where SAA fails closed, are logged separately.

SAA is licensed per device, not by usage or call volume. The evaluation is a paid, structured engagement; the fee is convertible into pilot credit or a strategic investment option after a successful go/no-go. Commercial licensing terms are discussed per deployment after the evaluation closes. Book a call to scope the evaluation and discuss terms.

Yes. Every routing decision is logged per device with a timestamp, producing a record of which speaker was classified as the principal speaker at each interaction point. For operations and compliance teams this is the attribution record a conventional voice stack does not produce: a structured log showing which interactions were addressed to the device and which were suppressed as PA, ambient, companion, or uncertain. The log is available for review during the evaluation and is part of the data delivered in the written go/no-go report.