Work

Support System for the Diagnosis of Vocal Pathologies

Year: 2025
Type: dissertation
Abstract: Diagnosis of voice disorders usually involves invasive techniques and requires coordinated action by multidisciplinary teams.The application of an automatic method based on speech processing offers a non-invasive and rapid solution suitable for both screening and preliminary diagnosis prior to evaluation by specialists.This thesis proposes a system to aid in the diagnosis of speech pathologies associated with the larynx.This system provides features such as jitter, shimmer, Harmonic to Noise Ratio (HNR) and autocorrelation and allows a diagnosis between healthy/pathological, with the possibility of obtaining the classification of the pathological group.The methods used involve classifiers typically used in the classification between healthy and pathological, as well as in the classification of pathologies or pathological groups, such as Multi Layer Perceptron (MLP) Neural Networks, Deep Learning models, namely Convolutional Neural Network (CNN).Features that model the source of the vocal tract, such as jitter, shimmer, HNR and Fundamental Frequency (F0), and features that model the vocal tract, such as Mel-Frequency Cepstral Coefficients (MFCC) and Line Spectral Frequency (LSF), were used.The databases used were the Saarbrucken Voice Database (SVD), Advanced Voice Function Assessment Databases (AVFAD) and the University of São Paulo (USP).During this work, a system to support the diagnosis of vocal pathologies was developed.This system was implemented at the Unidade Local de Saúde de São João (ULSSJ), where it was made available to medical staff, Dr Helana Vilarinho, Dr. João Pinto and Dra.Carla Moura, for testing.During the use by the doctors, a database was collected.In this approach, tests were carried out in order to classify healthy and pathological cases, and within the pathological ones to classify between 3 etiologies of pathologies (functional, organic and psychogenic) and within the organic ones to distinguish between structural and neurological.The recognition rate for healthy/pathological was 98.9% using a CNN.In the distinction between the 3 etiological groups, it was 89.1%, and in the distinction between the two organic groups, it was 91.9%, using Deep Neural Network (DNN) as a classifier.
Language: English
Open Access status: green
OpenAlex ID: W7124819183
Cites: 177
Cited by:
Related to: 10