Research On Feature Extraction And Classification Of Pathological Speech

Pathological speech usually refers to the speech in the condition that the voicesystem is damaged owing to disease. The vocal organs’lesions will change thecharacteristics of speech signal. As the method of analyzing speech signal to diagnoselesions is noninvasive, objectivity, convenience, the diagnosis and analysis based onpathological speech has become a current research hot spot. However, Because of thecomplexity of the voice system structure, the source of the disease is very difficult todetermine, so the research on analysis of pathological speech feature and diseasecharacteristics ‘correspondence is still in exploring stage.This article is focus on theresearch of pathological speech diagnostic methods. The main tasks are as follows:1.The traditional acoustic features of pathological speech are analyzed. Thetraditional acoustic features of pathologic speech is extracted to construct BasicAcoustics Feature Set(BAFS), including430-dimensional features.2. AS traditional acoustic features are extracted under the hypothesis that thespeech signal is short-time steady,which can’t reflect the dynamic change ofnon-stationary signal and S transform can better reflect the local characteristics of thenon-stationary signal, this paper proposes a new feature based on S transform.Compared with MFCC, it has good ability to express the pathology.3. As the traditional acoustic analysis are based on linear perspective of speechsignal, ignoring its nonlinear characteristics. This article analyzes the pathologyspeech from the nonlinear perspective and extract the three kind of nonlinear featuresof pathological speech as a supplement to the acoustic characteristics to construct arelatively complete set of pathological speech features.4. For the high dimensionality problem of the feature set，this paper proposes anew feature fusion and dimension reduction method based on visualizationtechnology.The method utilizes the characteristics of visual feature that it can be a goodexpression of the structural between the features and the characteristic of F-Score that itis capable of evaluating features ‘importance to achieve the goal of feature fusion anddimensionality reduction. With the particle swarm algorithm are compared, shows itsefficiency. Compared with the feature selection algorithm based on particle swarmoptimization, it is more efficient.5. Based on the above research, this paper established the pathological speechdiagnosis system, including the diagnosis module, the register module, the medicamentmanagement module, the teaching module and help module.The diagnosis moduleincluding feature extraction, feature dimension reduction and recognition classification of pathological speech. This system can provide clinical auxiliary diagnosis ofpathological speech.