Efficient Algorithm for Voice Disorders Identification Based on Support Vector Machines
Nawel Souissi, Adnane Chérif · Transylvanian Review · 2016
This paper proposes an efficient algorithm for accurate identification of pathological voices. In this context, a full analysis of the Mel Frequency Cepstral Coefficients (MFCC) was conducted; the Shannon entropy and energy were calculated directly from the time-domain signal of a frame. The efficiency of the energy and entropy features is well investigated and compared. A projection based Linear Discriminant Analysis (LDA) as feature reduction method is proposed in order to enhance the discriminative ability of the algorithm. Support Vector Machines (SVM) is suggested as classifier and the Sequential Minimal Optimization (SMO) as an optimisation algorithm was integrated with the SVM in order to improve the classification performances. Three hybrid combinations based on feature extraction methods are proposed and compared. The strength of the proposed system is evaluated based on two validation schemes; 5-fold cross validation and traditional validation. In addition, performance measures such as accuracy, sensitivity, specificity, precision and Area Under Curve (AUC) are investigated for every hybrid combination in order to establish a full comparative study and conclude the best hybrid algorithm that leads to the optimal recognition rates. The experimental results achieved with 120 disordered voice samples selected from ‘Saarbrucken Voice Database’ by the Institute of Phonetics of the University of the Saarland in Germany, show that the best optimized hybrid algorithm provides the highest accuracy rate of 100% and AUC of 100% with the 5-fold cross-validation.