Análise acústica, baseada no modelo linear de produção da fala, para discriminação de vozes patológicas.

Abstract

Pathological voice discrimination has been done by means of digital signal processing techniques as a complementary tool to videolaryngoscopic exams. This method is non-invasive to patients and more comfortable when compared to laryngoscopy. This work aims at the acoustic analysis of voice signals affected by laryngeal pathologies, particularly, vocal fold edema. The discrimination process of the pathological voice and consequently the pathology detection consists, basically, in three main stages: acoustic characterization, feature modeling and classification. The pathology is characterized by linear prediction, cepstral and mel-cepstral analysis. To estimate cepstral coefficients, a parametric approach derived from linear prediction analysis is used. The mel-cepstral coefficients estimation uses a nonparametric approach based on Fast Fourier Transform. An individual classifier is applied to each acoustic feature obtained to best evaluate its relevance in detecting the pathology presence. In order to reduce the amount of data related to the parameter vectors used in the analysis, a Vector Quantization technique is applied and a distortion measurement is associated to a preliminary stage of the classification process. For the final classification a parameter modeling is carried out using Hidden Markov Models as a refinement stage of the preliminary classification process. Results show that the developed methods are efficient in modeling the effects caused by the pathology in study and provide an efficient discrimination of the pathology when compared to normal voices.