Article
Siren Sounds as Acoustic Landmarks for Content Verification (en)
* Presenting author
Abstract:
In this study, we investigate the potential of siren sounds as acoustic landmarks for content verification, specifically focusing on its use to verify the regional origin of ambient audio recordings. Recognizing that sirens—whether from ambulances, firefighters, or police vehicles—are not only prevalent in urban environments but also may exhibit different sound characteristics, this research aims for siren sound detection and classification based on country of origin and siren type. To achieve this, two comprehensive datasets comprising recordings from nine countries and three siren types were compiled. We compare two approaches for siren sound classification: conventional supervised learning using deep learning, and a classification scheme based on proximity in a high-dimensional embedding space, which was derived from a pre-trained computer vision model. While originally being trained on object detection in natural images, we found that the latter model can discern subtle differences in the characteristic pitch contours of siren sounds. Our results confirm that different types of sirens can be well recognized and distinguished, but also that some siren types are used across national borders and therefore cannot be clearly assigned to a specific region.