Performance improvement of automatic pathological voice quality assessment based on higher-order statistics

Abstract

This thesis presents new parameters based on the HOS (Higher-Order Statistics) analysis to improve the classification performance of a multi-stage pathological voice assessment system. Automatic pathological diagnosis is a field which still demands further investigation mainly due to the difficulty in quantifying or standardizing the speech pathologists`` diagnoses. In recent years, various speech signal processing techniques have been proposed and applied for the voice disorder diagnosis. The objective is to quantitatively measure the degree of deviation of the pathological from the normal voice patterns with some acoustic analyses. And, objective supports of the diagnostics have some advantages to be adopted directly into the everyday life rather easily with less cost. Although most of the previous researches made novel contributions to the automatic detection of voice disorders and to voice quality assessment, their achievements are not easy to be compared with each other due to the lack of uniformity. Therefore, it is indispensable to compare the various pattern recognition techniques using a rather authorized disordered voice database. To comply with these necessities, we develop several pattern recognition algorithms which are more efficient and eligible to implement the system and propose some new parameters to improve the classification performance. They are the means, the variances, and the variations of the HOS such as the skewness and the kurtosis. Recently, the applications of the HOS to speech processing have been mainly motivated by the properties of Gaussian suppression and phase preservation. Works in this area are focused on the assumption that the HOS properties of speech are different from those of Gaussian noises. The proposed HOS-based parameters show meaningful differences among normal and pathological voices classified in the GRBAS scale. By employing these new parameters we design and implement the algorithm to classify pathologica...