Abnormal sound events detection using two-pass sound indexing method

Abstract

This paper proposes a two-pass sound indexing method for the detection of abnormal sound events, and applies it to an audio surveillance system. The proposed method extracts sound feature parameters in one-pass, constructs a sound source classification and feature parameters pool based on it, reduces the pool to one-pass detection candidate categories out of two-pass, reconstructs recognition feature vectors by concatenating the extracted feature parameters with single Mel-Frequency Cepstral Coefficient (MFCC) feature parameters, and performs abnormal sound detection using a Gaussian mixture models (GMM) algorithm. We verified the proposed method, by a performance comparison with single MFCC through simulation, showed excellent performance, and implemented abnormal sound events detection system on the basis of the result. ? 2014 International Information Institute.