Abstract
Automatic voice pathology diagnosis is a widely investigated area by the research community. Recently, in the literature, most of the proposed solutions are based on robust feature descriptors, which are combined with machine learning algorithms. Despite of their success, it is practically difficult to design handcrafted features which are optimal for specific classification tasks. Nowadays, deep learning approaches, particularly deep Convolutional Neural Networks (CNNs), have significant breakthroughs in the recognition tasks. In this study, the deep CNN, which was mainly explored in image recognition purposes, is used for the purpose of speech recognition. An approach is proposed for voice pathology recognition using both deep CNN and Genetic Algorithm (GA). The CNN weights are initialized using the solutions produced by GA, which minimizes the classification error and increases the ability to discriminate the voice pathology. Moreover, three popular deep CNN architectures, which have been investigated in the literature for image recognition, are adapted for voice pathology diagnosis, namely: AlexNet, VGG16, and ResNet34. For comparison purposes, performance of the hybrid CNN-GA algorithm is compared to the performance of the conventional CNN, and to some other approaches based on hybridization of deep CNN and meta-heuristic methods. Experimental results reveal that the improvement in voice pathology classification accuracy for proposed method in comparison to the basic CNN was 5.4% and when compared with other meta-heuristic based algorithms was up to 4.27%. The proposed approach also outperforms the state of the art works on the same dataset with overall accuracy of 99.37%.