Selection of pertinent acoustic features for detection of pathological voices

Abstract

This paper suggests a new method to improve the performance of acoustic features selection for the classification of pathological and normal voices. The effectiveness of the Mel Frequency Cepstrum Coefficients (MFCCs) using the Fisher Discriminant Ratio (FDR) is analyzed. To evaluate the performance of the selected features, experiments were performed using a Multi-Layer Perceptron (MLP) classifier with Feed Forward Back Propagation training algorithm (FFBP). The developed method was evaluated using voice data base composed of recorded voice samples (continuous speech) from normophonic and dysphonic speakers. Based on mixed voices database, the best selected features achieved a correct classification rate of 92.74%. The proposed system shows that the FDR is sufficiently a selection method of acoustic features for classification of pathological and normal voices.