Scream and gunshot detection in noisy environments

This paper describes an audio event detection system which automatically classifies an audio event as ambient noise, scream or gunshot. The classification system uses two parallel GMM classifiers for discriminating screams from noise and gunshots from noise. Each classifier is trained using different features, appropriately chosen from a set of 47 audio features, which are selected according to a 2-step process. First, feature subsets of increasing size are assembled using filter selection heuristics. Then, a classifier is trained and tested with each feature subset. The obtained classification performance is used to determine the optimal feature vector dimension. This allows a noticeable speed-up w.r.t. wrapper feature selection methods. In order to validate the proposed detection algorithm, we carried out extensive experiments on a rich set of gunshots and screams mixed with ambient noise at different SNRs. Our results demonstrate that the system is able to guarantee a precision of 90% at a false rejection rate of 8%.