Evaluation of discriminative features extracted from sustained vowel to classify normal and pathological speech

In this research ten features from sustained /a/ vowel have been extracted to achieve a method for automatic detection of voice disorders. These features include average pitch, highest and lowest pitch value in the signal, standard deviation of pitch, short time perturbation of pitch, standard deviation of amplitude coefficients, short time perturbation of amplitude coefficients, first formant, first to second formant power ratio, short time fractal dimension. In order to evaluate the capability of each feature to discriminate between normal and pathologic classes, between classes to inter classes variance ratio is calculated for each of ten features. Results proof the high ability of fractal feature to discriminate the classes. Then by a feed forward multi layer perceptron neural network the signals are classified using best three features. The classification procedure is independent of gender. The classification accuracy is 85.09% for test dataset.