Pattern recognition with applications to pre-diagnosis of pathologies in the vocal tract

摘要

For the detection of laryngeal pathologies, in general medical examinations, for example laryngoscopy and stroboscopy, are adopted. Besides being considered invasive and uncomfortable procedures, they are made only by medical request when the diseases are already on advanced levels. In order to perform a computational pre-diagnosis of such conditions, this paper presents a non-invasive technique in which three classifiers are tested and compared: Euclidian distance, RBF Neural Network with the Gaussian kernel, and RBF Neural Network with the modified Gaussian kernel. Based on a database of normal and pathological voices, tests that demonstrate the effectiveness of the proposed technique, which can be implemented in real-time, were performed.