Classification of Normal and Pathological Voices Using Convolutional Neural Network
Changwei Zhou, Lili Zhang, Zhang Xiaojun, Yuanbo Wu, Di Wu, Zhi Tao
Abstract
In this paper, the spectrogram features of voice are studied, and the differences between normal voice and pathological voice are analyzed. 53 cases of normal voice and 133 cases of pathological voice are selected from Massachusetts Eye and Ear Infirmary(MEEI) database for experiments. As the popularity of deep learning is increasing, convolutional neural network(CNN) based on tensorflow2.1 is applied to extract features from spectrograms., and the voice data are augmented for overfitting of the trained model. The recognition accuracy of pathological voice detection is 96.51%. The experimental results show that the characteristic parameters extracted from spectrograms by CNN have distinguish ability.