Multimodal Voice Pathology Detection With Voice Signal and Integrated Electroglottographic (EGG) Signal From the Saarbrücken Voice Database.

A powerful tool for communication, emotion, and artistic expression, the human voice plays a vital role in human interaction. However, approximately 25% of individuals globally experience voice disorders, often due to occupational strain, leading to significant challenges in communication and quality of life. This work proposes an automated, non-invasive system for detecting voice pathologies using voice and electroglottographic signals. The analysis focuses on high-pitched /a/ vowel sounds and spoken phrases from the Saarbrücken Voice Database, comprising both healthy and pathological samples recorded at varying pitch levels. Time, frequency, time-frequency domain, and cepstral features were employed to derive a variety of traits. Finding the most relevant features through feature selection techniques enhanced the performance of classification models. To categorize voice signals that are either healthy or pathological, several machine learning methods were used, including support vector machine (SVM), random forest (RF), decision tree (DT), and K-nearest neighbor (KNN). The proposed system offers a rapid, objective alternative to traditional diagnostic methods such as laryngoscopy, improving patient comfort and clinical accessibility. Model performance was evaluated using criteria such as receiver operating characteristic area, sensitivity, specificity, and accuracy. The classifiers achieved accuracies of 88% (SVM), 89% (RF), 82% (DT), and 86% (KNN), effectively detecting disorders such as laryngitis, leukoplakia, and spasmodic dysphonia. This work shows great promise for real-time clinical applications by fusing machine learning and sophisticated signal processing. This work emphasizes the importance of comparative evaluation and multimodal analysis in creating trustworthy vocal pathology detection systems.