Smart Data Driven System for Pathological Voices Classification

• Autor(a) principal: Fernandes, Joana
• Data de Publicação: 2022
• Outros Autores: Junior, Arnaldo Candido [UNESP], Freitas, Diamantino, Teixeira, João Paulo
• Tipo de documento: Artigo de conferência
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1007/978-3-031-23236-7_29; http://hdl.handle.net/11449/246824
• topic: Machine learning; Principal component analysis; Speech features; Speech pathologies; Vocal acoustic analysis

Classifying and recognizing voice pathologies non-invasively using acoustic analysis saves patient and specialist time and can improve the accuracy of assessments. In this work, we intend to understand which models provide better accuracy rates in the distinction between healthy and pathological, to later be implemented in a system for the detection of vocal pathologies. 194 control subjects and 350 pathological subjects distributed across 17 pathologies were used. Each subject has 3 vowels in 3 tones, which is equivalent to 9 sound files per subject. For each sound file, 13 parameters were extracted (jitta, jitter, Rap, PPQ5, ShdB, Shim, APQ3, APQ5, F0, HNR, autocorrelation, Shannon entropy and logarithmic entropy). For the classification between healthy and pathological, several classifiers were used (Decision Trees, Discriminant Analysis, Logistic Regression Classifiers, Naive Bayes Classifiers, Support Vector Machines, Nearest Neighbor Classifiers, Ensemble Classifiers, Neural Network Classifiers) with various models. For each patient, 118 parameters were used (13 acoustic parameters * 9 sound files per subject, plus the subject’s gender). As pre-processing of the input matrix data, the Outliers treatment was used using the quartile method, then the data were normalized and, finally, Principal Component Analysis (PCA) was applied in order to reduce the dimension. As the best model, the Wide Neural Network was obtained, with an accuracy of 98% and AUC of 0.99.