Detecção de desvios vocais utilizando modelos auto regressivos e o algoritmo KNN

Abstract

Some fields in Science propose to study vocal tract disorders from an analysis about voice vibration patterns. Generally, the weight of those researches is given by the identification – in a more specific level – of diseases in different stages of severity, which would be redressed through voice therapy or means that require more attention, hence generating the need of surgical procedures for its control. Although there are evidences in literature that the Digital Signal Processing allows a non-invasive diagnosis of laryngeal pathologies, such as vocal cord disorders, which provoke swelling, nodules, and paralyses, there is no definition of any most indicated method, and characteristics or appropriated parameters to detect voice deviations. Thus, the present paper proposes an algorithm to detect vocal deviances through the Voice Signal Analysis. In order to complete this study, it had been used data from the Disordered Voice Database, developed by the Massachusetts Eye and Ear Infirmary (MEEI) due to their wide use in researches regarding the voice and speech. A total of 166 signals from this database were used, including healthy voices and pathologic voices affected by swelling, nodule, and vocal fold paralysis. From the voice signals, autoregressive processes of order (AR and ARMA) were generated for a representation of those signals, and – by using the models’ parameters obtained – it had been used the KNN algorithm for a classification of the signals analyzed. Seeking an analysis of the efficiency of the algorithm proposed in this study, the results obtained from this algorithm were compared to a detection method, which only considers the Euclidian distance between the signals. The results found point that the propositioned method in this work presents a satisfactory result, generating a hit rate on the classification above 71% (more than the 31% from the use of the Euclidian distance). Moreover, the method used is easy to implement, so that it can be used along with simpler hardware. Consequently, this research has the potential to generate a cheap and accessible sorter for wide-scale use by health care professionals as a non-invasive pre-analysis to detect otorhinolaryngological pathologies that affect the voice.