Global statistical features-based approach for Acoustic Event Detection

Abstract

The analysis of acoustic data typically discusses the problem of segmenting the acoustic events into non-overlapping acoustically compact categories. In Acoustic Event Detection (AED), an acoustic event is categorized into speech and non-speech events. Detection of non-speech sounds such as scream, gun shots, explosions, and glass break events is very helpful in acoustic surveillance, multimedia information retrieval, and acoustic forensic applications. In this paper, we propose global statistical features-based representation for multi-variate varying length acoustic data. A discriminative model-based classifier is then used to classify different acoustic events. The proposed representation is of very less dimension. The proposed approach is evaluated on surveillance-oriented AED datasets such as CICESE (recorded from a smart room scenario), Environmental Sound Classification (ESC), and IEEE AASP/DCASE2013 (Office environment) datasets. The proposed approach gives a better performance when compared with the conventional Hidden Markov Model (HMM) and Gaussian Mixture Model (GMM) approaches.