关键词: 声音信号, 分类, 类所属, 特征选择, Mel频率倒谱系数（MFCC）

Abstract: Identifying speech and acoustic sounds is beneficial to the developments of some systems like security monitoring, healthcare and modern video-audio conferencing systems, etc. Since many of the acoustic signals have their own specific generation mechanism, there lacks of a systematic methodology on feature extraction for acoustic sounds. Considering the fact that different audio signals have their own specific characteristics, in the paper, the idea of class dependent feature selection and classification is utilized to solve the problem. Some commonly used acoustic features is studied and then a specific feature subset for each kind of sound signals is selected. With the selected class-dependent feature subsets, the support vector machine with radial basis function kernel is used as the classifier. The experiments are conducted on speech and two kinds of acoustic sounds, i.e., cough and the striking of cup-plate. Compared with the conventional feature selection method, class-dependent feature selection utilizes fewer features to realize a better classification.

Key words: acoustic sounds, classification, class-dependent, feature selection, Mel Frequency Cepstrum Coefficients（MFCC）