Passive acoustic monitoring with real-time AI classification of 14 threat categories for perimeter defence — entirely on-device.
At a glance
Built with
Overview
A defence installation needed to monitor acoustic signals across a wide perimeter and classify threats — vehicles, footsteps, digging, gunfire — in real time, without visual detection.
The Challenge
Acoustic classification in noisy outdoor environments is an unsolved hard problem. The system needed >92% accuracy at SNR as low as −5dB, with all processing on-device — no network latency tolerable.
How it fits together
Architecture
The Solution
We developed a CNN-Transformer hybrid trained on 400 hours of labelled site audio, adapting Whisper-style architecture for non-speech classification. It runs on edge nodes at sub-50ms inference across 14 threat categories with confidence scoring.
Results
From brief to production