On the use of electroglottography and speech signals for automatic classification of patients with voice pathologies

Abstract

ABSTRACT : Voice production is a crucial aspect of human life; problems with the voice can affect the quality of life by influencing how we communicate. Speech production involves various muscles and neural connections so that voice pathologies can arise from multiple sources. Early detection of these disorders is critical to maintaining or improving the patient's condition. However, diagnosing these pathologies is often time-consuming and subject to physicians' assessment. The increased popularity of Artificial Intelligence (AI) has led to the creation of machine learning and deep learning models that perform an automatic analysis based on patterns found in the data. These AI techniques offer the potential to simplify the diagnostic process, providing more consistent and objective assessments. However, they require a previous analysis of the data, the features that will be extracted, and the classifiers. This work analyzes and compares multiple techniques to classify voice pathologies using the Saarbrücken Voice Database, a German database containing multiple voice pathologies and healthy controls performing different tasks. This work aims to apply and compare different machine learning and deep learning techniques to find the best classifier, considering an unbiased analysis for age and gender. Also, this work aims to showcase the capabilities of a novel feature set called phase plots, which represented glottal cycles as elliptical trajectories superimposed in a 2D plane. Additionally, the study explores the impact of incorporating complementary information through early and late fusion methods on the classification process. By integrating these techniques, the study aims to enhance the accuracy and robustness of voice pathology classification. The findings of this work highlight the potential of automated techniques in voice pathology detection.