Pathological Voice Recognition Research By SVM Weights Match Feature Parameters

Language is one of the most commonly used, convenient and direct means ofcommunication, which is peculiar to mankind. With the changes of human life style, socialinteraction is increasing; also the disease incidence rate has increased significantly. Thanksto the development of pattern recognition technology, automatic and nondestructivedetection of pathological voice comes true, and the detection result is objective. Study onpathological voice recognition to replace the traditional clinical subjective detectionbecomes a research hotspot.The pathology voice recognition mainly focuses on the extraction and optimization offeature parameters. According to the feature parameter extraction, in order to improve theaccuracy of the fundamental frequency based parameters, the pitch detection algorithm ispresented based on optimal pathological voice reconstruction. The algorithm introduces thewavelet reconstruction of pathological voice signals removing the higher harmoniccomponents, and the use of approximate entropy and the largest Lyapunov exponent selectsoptimal reconstructed pathological voices adaptively for fundamental frequency extraction.This algorithm can effectively suppress the harmonic and half-frequency errors caused bytraditional pitch detection algorithm. Furthermore, in this paper, the characteristicparameter wavelet energy-entropy-ratio is proposed in the nonlinear analysis. Thisparameter takes advantage of the wavelet energy and wavelet entropy, and wellcharacterized the pathological voice energy and complexity.In the aspect of feature parameters optimization, in order to reduce the dimension offeature parameters, feature parameters fusion is carried out using support vector machineweighted matching fusion algorithm. First, Spearman correlation analysis of the traditional characteristic parameters selects acoustic parameters of small relatively to form thetraditional parameter set. Then kernel principal component analysis is applied to reduce thedimension of cepstral feature parameters, and MFCC and LPCC reduce redundancy as well.The weighted matching of three parameter sets using support vector machine givesparameter sets appropriate weight, eventually combine into an optimal parameter set.In this paper, the experimental part first carries on the statistical analysis ofpathological voice database, in addition, various parameters are analyzed, and recognitionstudies on optimal parameter set of features by use of multiple classifier. Compared withthe original features, the rate of pathological voice recognition algorithm is up to96.92%with the normal voice, vocal nodules and vocal polyp’s recognition rate is up to83.22%.The recognition rates are improved.