부산대학교 도서관

Voice disorders affect a significant portion of the global population, particularly those in vocally demanding professions such as singers, actors, teachers, and lawyers. Early detection and diagnosis of voice pathology diseases are critical to improving treatment outcomes and preventing further damage to the vocal cords. Digital processing of speech signals has emerged as a promising technique for analyzing voice vibrations and identifying deformities in the vocal cord function. In this paper, a cost-effective computational method involves processing the voice signal by passing to a stack of band-pass filters, dividing the processed signal of each filter to a set of overlapped frames, applying the autocorrelation formula to every single frame, and using entropy to extract features. The method has shown promise in reliably detecting and classifying voice pathology diseases, but further research is required to confirm its efficacy and reliability. Deep learning algorithms and Mel spectrogram feature extraction techniques are present in this paper for voice pathology detection. VGG16, VGG19, and ResNet50 are compared. The system demonstrated high prediction accuracy results using ResNet50 on training and testing dataset. The system shows potential for clinical applications in voice disorder assessment and diagnosis. The system also holds promise as a telemedicine tool, enabling remote assessment and monitoring of patients' vocal health.