Detecção e classificação de categorias de disfonias com redes neurais convolucionais

Abstract

Conducted surveys worldwide show that 16.9% to 35.8% of the general population experience or claim to have experienced some degree of vocal disorder. However, the unavailability of trained professionals or diagnostic equipment, among other factors, may result in the non-treatment of patients and the consequent worsening of their quality of life. Recent advances in computational science have enabled the use of methodologies for automatically detecting dysphonia based on machine learning techniques to complement clinical evaluation. However, such methods only explore the binary distinction between healthy voices and dysphonia or perform a limited classification step using the most represented dysphonia types in the databases. Because of this, the present work evaluates a new methodology for classifying dysphonia based on its grouping into three categories: Functional Dysphonia, Organic Dysphonia, and Organofunctional Dysphonia. This grouping was applied to voice recordings of two databases: the Saarbruecken Voice Database (SVD) and the Advanced Voice Function Assessment Database (AVFAD). After this grouping, a feature extraction step was applied to the audio signals using spectrograms, followed by a classification step using convolutional neural networks. From the results obtained, it is valid to state that the method effectively detects organic and organofunctional dysphonia, reaching test accuracies of 76.1% and 72.2%, respectively, for the SVD and 82. 8% and 77.3% for AVFAD. However, it was impossible to successfully distinguish functional dysphonia, since they are underrepresented in the databases, which negatively impacted the overall performance of the classifier, which reached 53.2% for SVD, and 59.8% for AVFAD. However, an improvement of this methodology can increase the capacity to detect functional dysphonia, improving its performance.