A Wavelet Based MFCC Approach for the Phoneme Independent Pathological Voice Detection

Abstract

This paper proposes a new approach for the phoneme independent pathological voice detection. The phonemes /a/, /i/, /u/ from normal and subjects suffering from voice disorders are recorded. The system uses wavelet based Mel Frequency Cepstral Coefficients (MFCCs) as features, which are given to Gaussian Mixture Model-Universal Background Model (GMM-UBM) classifier. The MFCCs are computed for each wavelet sub band and GMM-UBM score is obtained. The decision is taken by combining GMM-UBM scores of individual sub bands. When the 18MFCC features are given to GMM-UBM classifier it can be seen that the accuracy is 85.18%. But when the wavelet based 18MFCCs are given, the accuracy is 93.32%, which indicates that wavelet based MFCCs improves the classification accuracy.