Automatic Classification For Pathological Voice Based On Acoustic Parameters And SVM

With the increase of human society communication and the change of habits and customs, the incidence of voice diseases is increasing. Therefore, the clinical assessment of speech voice disorders is emphasized, and there is also more research on pathological voice detection and analysis. The accurate assessment of voice disorders is fundamental to the solution of all voice problems. This study is prospective which has launched objective automatic classification and grade evaluation on pathological voice revolving around the acoustic measurements of pathological voice as well as Support Vector Machine in pattern recognition methods, according to the current research status and trends of pathological voice. The automatic classifiers of pathological voice are built eventually.The acoustic characteristic parameters of pathological voice have been extracted. All six types of parameters were extracted by a series of voice signal processing technology; that is, the formant and amplitude parameters were achieved by Linear Prediction, the glottal parameters were achieved by Inverse Filtering, and the fundamental frequency, HNR and cepstral parameters were achieved by Homomorphic Analysis and Cepstrum Technology. The pathological voice classifiers were built to need inputting sevaral feature vectors, which were constituted by the above acoustic parameters.The acoustic characteristic parameters of pathological voice have been analysized and optimized. The perceptual evaluation of pathological voice was achieved by GBRAS, and then the grade evaluation results of G, R, B were obtained. The optimized11parameters were obtained by correlation analysis between all parameters as well as between the parameters and grade evaluation results. These parameters that had low correlation with each other and high with the grade evaluation results were inputted to SVM.The pathological voice classifiers have been implemented. SVM was selected to classify the voice. The training models of SVM were built based on two types of acoustic characteristic vectors so that the classifiers were achieved for classification of normal and pathological voice. Then the classifiers were achieved for classification of pathological voice severity level by one-on-one approach on this basis.The effect of pathological voice classifiers have been verified by Cross Validation and ROC curve. The effect to classify pathological voice from normal was the best and the rate of recognition reached96%～98%, so it could realize the distinction between the above two types of voice. The effect to classify pathological voice severity level was certain and the rate of recognition reached73%～84%. The effect of classifiers based on the optimized parameters was slightly lower than that based on the initial parameters, so the former could be applied in the objective automatic classification on pathological voice for greater efficiency.The pathological voice classifiers can differentiate objectively not only normal and pathological voice, but also pathological voice severity level. The assessment of voice disorders will be more objective and not be affected by subjective differences, language environment and else factors through the research on objective classification and grade evaluation of pathological voice.