Dysarthria Severity Classification Using Phase Based Features of LP Residual

Classifying the severity of speech impairment due to dysarthria is crucial for optimizing care and enhancing communication abilities for affected individuals. This study explores the use of the Modified Group Delay Function (MGDF) of LP residual signal in classifying dysarthria severity-levels. Evaluations were conducted using standard UA-Speech and TORGO datasets. A stratified Convolutional Neural Network (CNN) with 5-fold cross-validation validated the results. Baseline features included Linear Frequency Cepstral Coefficients (LFCC), Mel Frequency Cepstral Coefficients (MFCC), and Whisper module. Key performance evaluation metrics were accuracy, precision, recall, and F1-score. Finally, the latency period was analyzed for practical deployment of the system, system’s ability to accurately recognize and process speech from any speaker, without needing to be specifically trained or adapted to individual voice characteristics.