ConVox: A Robust Deep Learning Approach for Accurate Voice Disorder Detection with Multilingual Capabilities

Abstract

Voice disorders significantly impact an individual's ability to communicate verbally, particularly affecting the elderly community. Diagnosing these disorders is complex, often hindered by the limitations of traditional imaging techniques. This study presents a novel deep learning framework for voice disorder detection through audio classification, addressing the challenge of diagnosing these disorders that affect a large proportion of elderly adults in America. Our model, ConVox, utilizes a sequential stack of one-dimensional convolutional neural networks to conduct binary classification of voice disorders. We leverage four large datasets: Advanced Voice Function Assessment Databases, Saarbrücken Voice Database, TORGO Database, and UA Speech Database, which together comprise 22,883 audio samples in Waveform Audio File Format. The model achieved notable accuracies of 99.89% in training, 99.91% in validation, and 99.74% in testing, outperforming existing models. With an area-under-curve of 0.999995, precision of 0.9972, and recall of 0.9994, our model demonstrates exceptional performance in accurately identifying voice disorders with a very low rate of false positives and false negatives. Additionally, this model demonstrates promising performance across multiple languages and voice pathologies. ConVox's comprehensiveness and high accuracy demonstrate that it is a promising tool for audio classification, potentially enhancing healthcare outcomes for individuals with voice disorders.