Automatic Classification of Disordered Voices Based on a Hybrid HMM-SVM Model
Redouane Benhammoud, Abdellah Kacha · Journal of Communications Technology and Electronics · 2021
Abstract The main concern of this paper is the clinical assessment of disordered voices using an automatic classification method based on a hybrid hidden Marcov model (HMM) and support vector machine (SVM). We investigate the effectiveness of the Mel-frequency cepstral coefficients (MFCC), log-frequency power coefficients (LFPC) and linear predictive cepstral coefficients (LPCC) as acoustic features for the classifier. The efficiency of the hybrid HMM-SVM model is tested on a concatenation of two Dutch sentences spoken by 28 normophonic speakers and 223 pathological speakers with several levels of dysphonia. The performance of the hybrid HMM-SVM classifier in terms of classification accuracy is compared with that of the conventional HMM and SVM classifiers when used separately. The highest two and three categories classification accuracies obtained by the hybrid HMM-SVM classifier are 97.35 and 92.01%, respectively, while the highest classification accuracies obtained by the two and three class classification with the HMMs are respectively 89.16 and 80.69%. The highest two and three class classification accuracies obtained by the SVM classifier with a radial basis function (RBF) kernel are respectively 77.36 and 75.15%, while the results obtained by the SVM classifier with a linear kernel are respectively 75.22 and 71.33% for the classification into two and three categories.