How the Army's Warden 2026 Event Advanced Certus’s TEAM ML Developments Towards Operationalization
UAS Strategic’s Zev Nadler recently supported the team at Certus Innovations in collecting over 3.5 hours of audio from the Army's Warden 2026 event.
While the data was live and operationally realistic, a military exercise with no gunshots was a great opportunity to test DART (Distributed Acoustic Reasoner for TAK), Certus’s TEAM-ML TAK plugin that provides real-time gunshot detection, classification and localization. In June, DART achieved 100% precision (no false alarms) on this acoustic data, positioning it solidly on the path to operationalization.
To test DART’s capabilities further, the Warden 2026 event dataset was then combined with unlabeled street noise from the TUT Sound Events 2017 dataset in order to “smoke test” the model on a 5-hour nontrivial, realistic, and unseen setting. Even on this more difficult task, DART scored 100% precision.
Next, the audio from the Army's Warden 2026 event also included unlabeled UAS recordings at unknown times. Extracting each instance of sUAS audio from the 12,127 total seconds of audio recording is complex. However, FALCON, Certus’s TEAM-ML TAK plugin that provides UAS detection, classification, and localization, was able to prefilter it, extracting each of the one second clips of UAS audio.
Certus’s FALCON detector is trained to detect sUAS down to a signal-to-noise ratio of -3dB, which is about twice as much noise as signal. While a human ear can’t always pick out a UAS that quiet, the model detects sUAS at least down to the limit of human hearing, if not further.
After annotating the Warden dataset and testing FALCON, the model achieved 99.3% accuracy, not quite field ready, but a positive sign for its first real field test.
