Reconstructed Phase Space and Convolutional Neural Networks for Classifying Voice Pathologies

In this paper, we present a new method for classifying voice pathologies. Reconstructed Phase Space (RPS) images are employed to represent the nonlinear dynamics of the signals, and a Convolutional Neural Network (CNN) is designed to automatically learn spatial features and a classification decision from the RPS images. Due to the large parameter space of the CNN, we augmented the Massachusetts Eye and Ear Infirmary (MEEI) database with synthetic training data obtained by slowing down or speeding up the audio signal. The proposed method was evaluated in the pairwise classification of 5 voice pathologies: paralysis, edema, nodule, polyp and keratosis. Experiments were also carried out on a broader pathology class, called benign lesion, consisting of nodule, polyp and cyst signals. Accuracies similar to state-of-the-art approaches support the relevance of the method. Best accuracy was achieved in the polyp vs. nodule classification. Data augmentation was beneficial to most of the classification experiments.