Discrimination of environmental background noise in presence of speech using sample-pairs statistics based features

摘要

A methodology to discriminate the different classes of background noise using new features based on samples of the signal is presented here. Two consecutive samples of different amplitude of the discretetime signals are termed as sample-pair and 14 types of sample-pairs are considered here as fundamental features. Results of simulation work proves that count of some of such type of sample-pairs as well as count of few combinations of two, three and four such sample-pairs are useful to detect and discriminate the different acoustic noise mixed with speech signals. On the basis of simulation results, the performance of proposed features have proved better than other spectral features like Mel Frequency Cepstral Coefficients (MFCC), Spectral Centroid, Spectral Flux and Spectral Roll-off regarding discrimination capabilities, simplicity of extraction process and lesser dependency over speech utterances mixed with noise. These sample-pairs based features having advantage of not requiring frame-decomposition and silence period removal. Their discrimination capabilities are shown by Fisher's F-ratio as performance index. The multiclass Support Vector Machine (SVM) is used as a classifier.