Classification of normal and pathological infant cries using bispectrum features

Abstract

In this paper, bispectrum-based feature extraction method is proposed for classification of normal vs. pathological infant cries. Bispectrum is computed for all segments of normal as well as pathological cries. Bispectrum is a two-dimensional (2-D) feature. A tensor is formed using these bispectrum features and then for feature reduction, higher order singular value decomposition theorem (HOSVD) is applied. Our experimental results show 70.56 % average accuracy of classification with support vector machine (SVM) classifier, whereas baseline features, viz., MFCC, LPC and PLP gave classification accuracy of 52.41 %, 61.27 % and 57.41 %, respectively. For showing the effectiveness of the proposed feature extraction method, a comparison with other feature extraction methods which uses diagonal slice and peaks and their locations as feature vectors is given as well.