Pathological Voice Detection Using Efficient Combination of Heterogeneous Features

Abstract

SUMMARY Combination of mutually complementary features is necessary to cope with various changes in pattern classification between normal and pathological voices. This paper proposes a method to improve pathological/normal voice classification performance by combining heterogeneous features. Different combinations of auditory-based and higherorder features are investigated. Their performances are measured by Gaussian mixture models (GMMs), linear discriminant analysis (LDA), and a classification and regression tree (CART) method. The proposed classification method by using the CART analysis is shown to be an effective method for pathological voice detection, with a 92.7% classification performance rate. This is a noticeable improvement of 54.32% compared to the MFCC-based GMM algorithm in terms of error reduction. key words: pathological voice detection, heterogeneous feature combination, mel frequency filter bank energies, higher-order statistics, pattern classification algorithm

Feature Extraction of Elderly Signals based on Bicoherence Estimation for Automated Medical Diagnosis System

“…If it is a little larger deviation from an average value, it may have an effect on the performance degradation of voice analysis and recognition [14]- [15]. In case of most welfare smart devices, the elderly signals have been neglected due to interface which does not take into account the elderly [6].…”

“…Further, the combination of HOS analysis and the linear predictive coding (LPC) residual may help to effectively construct important information to distinguish the signal types of disordered voices [17]- [18]. Although HOS analysis holds promise as one possible method of distinguishing between normal/pathological voices and signal type classification [14]- [16], no studies have applied HOS analysis to analyze elderly voices. Therefore, novel HOS-based parameters estimated in frequency domain for analysis of elderly voices are proposed in this paper.…”

An efficient approach using HOS-based parameters in the LPC residual domain to classify breathy and rough voices

“…While previous work in the area of speech analysis, voicing classification, and pitch estimation have attempted to exploit parameters based on the HOSs of speech signals, little has been done in providing an analytical framework for disordered voice analysis [20][21][22][23][24][25]. Alonso et al proposed the following HOS-based parameters: the interference value of the bicoherence index, the relative values of the high and low frequency energy of the onedimensional bicoherence index, the value of their interference, the variation of the estimated noise by means of the module of the bispectrum, the variation of the noise obtained through the same technique taking into account in this case the phase information, and the interference of the kurtosis [23].…”

“…We proposed a method to improve the classification performance between pathological and normal voices by combining heterogeneous parameters like auditory-based parameters, the normalized skewness, and the normalized kurtosis [24]. The performances of these measures were measured by Gaussian mixture models (GMMs), linear discriminant analysis (LDA), and CART methods.…”

“…CART can be used to effectively combine multiple parameters [24,25,30,31]. Especially, compared with other methods like HMM, GMM, and LDA, it is powerful because it can deal with incomplete and small size data.…”