The classification of pathological voice from healthy voice was studied based upon 27 acoustic features derived from a single sound signal of vowel /a:/. First, the feature space was transferred to reduce the data dim...The classification of pathological voice from healthy voice was studied based upon 27 acoustic features derived from a single sound signal of vowel /a:/. First, the feature space was transferred to reduce the data dimension by principle component analysis (PCA). Then the voice samples were classified according to the reduced PCA parameters by support vector machine (SVM) using radial basis function (RBF) as a kernel function. Meanwhile, by changing the ratio of opposite class samples, the accuracy under different features combinations was tested. Experimental data were provided by the voice database of Massachusetts Eye and Ear Infirmary (MEEI) in which 216 vowel /a:/ samples were collected from subjects of healthy and pathological cases, and tested with 5 fold cross-validation method. The result shows the positive rate of pathological voices was improved from 92% to 98% through the PCA method. STD, Fatr, Tasm, NHR, SEG, and PER are pathology sensitive features in illness detection. Using these sensitive features the accuracy of detection of pathological voice from healthy voice can reach 97%.展开更多
文摘The classification of pathological voice from healthy voice was studied based upon 27 acoustic features derived from a single sound signal of vowel /a:/. First, the feature space was transferred to reduce the data dimension by principle component analysis (PCA). Then the voice samples were classified according to the reduced PCA parameters by support vector machine (SVM) using radial basis function (RBF) as a kernel function. Meanwhile, by changing the ratio of opposite class samples, the accuracy under different features combinations was tested. Experimental data were provided by the voice database of Massachusetts Eye and Ear Infirmary (MEEI) in which 216 vowel /a:/ samples were collected from subjects of healthy and pathological cases, and tested with 5 fold cross-validation method. The result shows the positive rate of pathological voices was improved from 92% to 98% through the PCA method. STD, Fatr, Tasm, NHR, SEG, and PER are pathology sensitive features in illness detection. Using these sensitive features the accuracy of detection of pathological voice from healthy voice can reach 97%.
