Yet Another Approach for the Measurement of the Degree of Voice Normality: A Simple Scheme Based on Feature Reduction and Single Gaussian Distributions

摘要

In this paper, we propose another approach for the measurement of the degree of voice normality based on statistical modeling. The basic methodology behind the proposed approach is the "Pathological Likelihood Index" reported by Godino-Llorente JI. et al. [1]. The major innovations are: exploring a reduced set of Mel frequency cepstral coefficients (MFCC) and ignoring their derivatives, a linear projection of MFCCs into one dimensional space using Fisher's linear discriminant, and, modeling build around single Gaussian distributions instead of mixtures of distributions. We have evaluated the proposed approach using the Massachusetts Eye and Ear Infirmary database (MEEI). The obtained results are better than the one reported in [1].