Study On Pathological Speech Detection And Classification Based On Acoustic And Kinematic Features

Due to the loss of hearing ability,patients with hearing impairment generally have the problem of dysarthria.For the detection and classification of pathological speech in hearing-impaired patients,the current research focuses on pathological speech acoustic features,pays less attention to other types of features,and has incomplete feature extraction problems.Based on the background significance and research status of pathological speech detection and classification,this paper summarizes the traditional feature extraction and pattern recognition methods of pathological speech detection and classification.To improve the result of pathological speech detection and classification,from the aspects of acoustics and kinematics,to analyze and improve the features of pathological speech,the specific research contents of the paper are as follows:1.This paper takes the pronunciation of hearing impaired patients as the research object,analyzes the pronunciation characteristics of patients,and elaborates on the feature extraction and pattern recognition in the pathological speech detection system.In the feature extraction section,the characteristics and extraction process of commonly used features of pathological speech are introduced.In the pattern recognition section,the principles of three classification models of support vector machine,random forest,and multilayer perceptron are introduced.Then this paper extracts the traditional acoustic features of the hearing impaired patients’ pronunciation,and combines the classification model to achieve pathological speech detection and classification.2.Aiming at the problem of incomplete speech feature extraction using time-frequency analysis method on the entire speech,this paper proposes a multi-scale autocorrelation feature based on empirical mode decomposition.First,this paper uses empirical mode decomposition technology to decompose the speech signal adaptively;then calculates the autocorrelation function of the signals in different frequency bands to extract the multi-scale autocorrelation features.Finally,in order to verify the validity of the features in this paper,the multi-scale autocorrelation features based on wavelet transform are extracted as comparison features,and the detection and classification results of the features in this paper,the multi-scale autocorrelation features based on wavelet decomposition and traditional acoustic features are compared3.In order to further improve the result of pathological speech detection and classification,this paper proposes a pathological speech detection and classification method based on acoustic and kinematic features.First,this paper analyzes the trajectory of the lips and tongue in the process of generating pathological speech and normal speech,extracts the kinematic features of the two parts,and compares the detection and classification results of the combined kinematic features of the lips,tongue,and two parts.Then,compare the detection and classification results of kinematics and acoustic features to verify the classification performance of kinematics features.Finally,the kinematics features and acoustic features are fused into features,and the kernel principal component analysis method is used to reduce dimensionality.In the Chinese letter and Chinese monosyllable corpus,we compare the detection and classification result of fused features,single features,and different fused features.By comparing the classification result of multi-scale autocorrelation features,traditional acoustic features,kinematics features,and different fusion features of pathological speech,it is concluded that the classification result of multi-scale autocorrelation features in this paper is superior to other acoustic features.After it is integrated with kinematic features,the classification result is significantly improved.The best pathological speech detection accuracy of Chinese letters and Chinese monosyllables reaches 94.5% and 96.3%,respectively,and the best classification accuracy reaches 95.7% and 96.8%respectively.This method provides a reference for the study of pathological speech automatic diagnosis technology for hearing impaired patients.