Milan Gupta

Diagnosis of Dysarthria by Machine Learning.

Started Feb. 1, 2024

Abstract or project description

Those with dysarthria have problems in the nerves and muscles controlling speech, which usually leads to unclear or extremely unclear spoken words. While many studies have been carried out to examine speech impairment, the variation of this problem among people with a similar dysarthria diagnosis has necessitated the need for more research in this area. The particular type and severity of the impairment are essential in monitoring the progress of dysarthria and making effective therapeutic interventions. This project, therefore, describes a Convolutional Neural Network (CNN) model for dysarthria detection, where several acoustic features are extracted in the form of zero crossing rates, Mel Frequency Cepstral Coefficients (MFCCs), spectral centroids, and spectral roll-off. Using the TORGO database of speech signals, training the model, and testing it for its efficiency has shown much promise in the early diagnosis of dysarthric speech. The numerical results indicate that the model designed provides an efficiency of nearly 95%, which is higher than previous model architectures. This model aims to identify the condition early and help improve the management of dysarthria through timely and accurate diagnosis.