Automatic Conflict Detection in Police Body-Worn Audio

In this paper we propose a novel method for automatic conflict detection in police body-worn audio (BWA). Methodologies from statistics, signal processing and machine learning play a burgeoning role in criminology and predictive policing [2], but such tools have not yet been explored for conflict detection in body-worn recordings. Moreover, we find that existing approaches are ineffective when applied to these data off-the-shelf. Notable papers on conflict escalation investigate speech overlap (interruption) and conversational turn-taking as indicators of conflict in political debates. In [3], overlap statistics directly present in a hand-labelled dataset are used to predict conflict, while [4] detect overlap through a Support Vector Machine (SVM) with acoustic and prosodic features. The work in [5] compares variations on both methods. Using automatic overlap detection, their method achieves 62.3% unweighted conflict accuracy at best in political debate audio. This approach is all the less effective on BWA data, which is far noisier and more diverse.