Estudo de técnicas para classificação de vozes afetadas por patologias.

Abstract

In recent years, several studies in digital voice processing are being made in order to create techniques to support a noninvasive accurate diagnosis of vocal tract diseases by aspecialist, making the patient feel comfortable during examination. This work deals with the investigation of techniques for classification of voices affected by laryngeal pathologies, especially Reinke’s edema, aiming to build a support system to the specialist. The system for the diagnosis of laryngeal pathologies, proposed here, consists of three main steps: preprocessing the speech signal, feature extraction and classification. Preprocessing corresponds the acquisition of voice signal, the application of a pre-emphasis filter for minimizing the radiation effects from the lips and from variation in glottal area, and the signal segmentation and windowing. The non-use of pre-emphasis was also investigated at this point. In the feature extraction step, we use coefficients obtained from the linear prediction analysis (LPC coefficients), cepstral coefficients, delta-cepstral coefficients, and afeature vectorc ombining LPC and cepstral coefficients. The classification is divided into two parts: classification of normal voice versus voice affected by pathology, without specifying which pathology, and if the signal is classified as voice affected by pathology, second part happens, which is performed by the classification between voice affected by Reinke’s edema and voice affected by other pathology. For both parties, 3 different classifiers were tested: Neural Networks Multilayer Perceptron - MLP, Gaussian Mixture Models and Vector Quantization. To differentiate between normal voice and voice affected by pathology, the best results were obtained using Neural Networks. To differentiate between voice affected by edema and voice affected by pathology, the best results were obtained using vector quantization. In both cases, the best results were obtained when usingcepstral coefficients and withoutuse of pre-emphasis.