Deep learning aplicado a classificação de patologias da voz

Abstract

The classification of voice related pathologies using Deep Learning concepts has beenincreasing considerably in recent years. Good results have already been obtained forclassification in sustained speech with vowels, but there are still few studies related tothe classification of this problem using continuous speech. Therefore, the focus of thisdissertation is to implement the main models of Deep Learning for the classification ofvoice pathologies in continuous speech, using the German phrase "Guten Morgen, wiegeht es Ihnen?"From the Saarbruecken Voice Database. The pathologies of dysphonia,laryngitis and paralysis of the vocal cords, as well as the healthy class, are used formulti-class and binary analyzes. In addition, a previous study for the classification withvowels in the same pathologies is also carried out. The best result for the vowels is 99% accuracy for the implementation of an LSTM model with parameters Jitter, Shimmerand Autocorrelation, in the binary classification between laryngitis and healthy. Forthe phrases, a comparative study between neural networks, convolutional and recurrentmodels with the parameter MFCCs and Spectrograms in the Mel scale, obtaining resultsof 76% F-measure for dysphonia x healthy, 68% F-measure for laryngitis x healthy, 80%F-measure for healthy x paralysis of the vocal cords. For multi-class classification isobtained 59% and 40% of F-measure for 3 classes and 4 classes, respectively.