Evaluating RNN Variants for Dysphonia Classification Using the Uncommon Voice Dataset: A Comparative Analysis

Abstract: Dysphonia, a voice disorder caused by vocal cord dysfunction, significantly affects individuals' communication abilities. Early and accurate detection of dysphonia is crucial for timely intervention and effective treatment. Leveraging advancements in deep learning, this study employs RNN - Recurrent Neural Networks to enhance detection accuracy. However, most studies in this domain rely on common datasets, leading to limited generalizability of models to diverse populations. In this study, we explore the use of various RNN variants, including traditional RNN - Recurrent Neural Network, LSTM - Long Short-Term Memory and GRU - Gated Recurrent Neural Network, to detect dysphonia in an uncommon voice dataset. Existing works focus primarily on conventional datasets and simpler classifiers, leaving room for improvement in accuracy and robustness. Our methodology leverages feature extraction techniques to preprocess the dataset, followed by training RNN variants to evaluate their performance in classifying dysphonic and non-dysphonic voices. Each RNN variant was trained and evaluated on the preprocessed dataset, divided into an 80:20 ratio for training and testing. The results revealed differences in model performance, with the standard RNN achieving an accuracy of 76%, while the LSTM and GRU models outperformed it, achieving accuracy of 94% and 93%, respectively. The experimental results demonstrate the effectiveness of advanced RNN models in handling diverse and challenging datasets, offering insights into their comparative performance for dysphonia detection and advancing research in voice disorder diagnosis. Keywords: Dysphonia, Disorder, RNN, LSTM, GRU.