Context-Aware Acoustic Threat Detection Using a Partially Fine-Tuned Audio Spectrogram Transformer
Research Digest on Engineering Management and Social Innovations · 31.03.2026
DOI: 10.46647/icetetas017 ↗Kurz gefasst
Fachartikel in der Zeitschrift Research Digest on Engineering Management and Social Innovations, online erschienen am 31.03.2026. Verfasst von Y. C. Ashok Kumar und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Background
Deploying acoustic surveillance in real-world environments is harder than it might seem. Off-the-shelf sound classifiers trained on broad datasets recognise hundreds of event types reasonably well, but they are poorly calibrated for the small subset of sounds—gunshots, shattered glass, screams—that actually matter in a security context. Purpose-built threat detectors tend to be brittle outside their narrow training distribution. Getting both capabilities into one model at the same time has remained an open challenge.
Objective
In this paper we describe how we tackled this problem by partially fine-tuning a pretrained Audio Spectrogram Transformer (AST) originally trained on the full 527-class AudioSet ontology. Rather than retraining the whole network, we kept the first six transformer layers frozen—preserving the general-purpose acoustic representations—and only updated the upper six layers on a curated mix of threat and non-threat audio drawn from FSD50K, ESC-50, and UrbanSound8K.
Results
Our model, PFT-AST, reaches a Macro-F1 of 0.88, an AUC of 0.962, and a false alarm rate of just 3.2%—less than half the 7.1% recorded by the AST baseline. On a held-out test partition, 3,868 threat events were correctly flagged with only 63 misses.
Conclusion
Selectively re-adapting just the upper transformer layers is an effective, parameter-efficient way to inject threat awareness into a general-purpose audio model without dismantling what it already knows.
Zitieren
APA
Kumar, Y. C. A., & Ragi, R. (2026). Context-Aware Acoustic Threat Detection Using a Partially Fine-Tuned Audio Spectrogram Transformer. Research Digest on Engineering Management and Social Innovations. https://doi.org/10.46647/icetetas017
Vancouver
Kumar YCA, Ragi R. Context-Aware Acoustic Threat Detection Using a Partially Fine-Tuned Audio Spectrogram Transformer. Research Digest on Engineering Management and Social Innovations. 2026. doi:10.46647/icetetas017
BibTeX
@article{Kumar2026,
author = {Kumar, Y. C. Ashok and Ragi, Rajesh},
title = {Context-Aware Acoustic Threat Detection Using a Partially Fine-Tuned Audio Spectrogram Transformer},
journal = {Research Digest on Engineering Management and Social Innovations},
year = {2026},
doi = {10.46647/icetetas017}
}
EUVIMED