The Study of Voice Pathology Detection based on MFCC and SVM

Abstract

Subjective auditory perception evaluation of voice is the most simple and direct method for judgment of the degree of voice lesions and the treatment effect. But it is closely related to the clinical experience of doctors. Recently, some voice automatic diagnosis methods based on voice feature parameters and classification algorithms have been proposed. Mel Frequency Cepstral Coefficient (MFCC) is the most commonly used feature parameter. However, it is not clear the role of MFCC dynamic features in improving diagnosis results. This study adopted the features of MFCC, MFCC + ΔMFCC, and MFCC + ΔMFCC + ΔΔMFCC respectively, combined with the Support Vector Machine (SVM) method to further determine whether adding dynamic MFCC features can improve the accuracy of pathological voice detection. The results showed that no matter whether dynamic features were added or not, the accuracy rate and specificity have not changed significantly. This means the dynamic change of the MFCC characteristic parameters is slight at least for vowel vocalization. This study may provide useful information for pathological voice diagnosis based on vowel vocalization.