Dysarthria detection using convolution neural network

Abstract

Dysarthria patients have difficulty controlling their speaking muscles, resulting in incomprehensible speech. A number of studies have looked into speech impairments; however, more research is needed to consider speakers with the same impairment but different levels of impairment. The type of impairment and severity level will aid in determining the dysarthria's progression as well as treatment planning. The use of a Convolutional Neural Network-based model to detect dysarthria is proposed in this paper. Early detection is the first step toward better impairment management. For the analysis of speech signals, the proposed model uses a number of speech features such as zero crossing rates, MFCCs, spectral centroids, and spectral roll off. For training and testing, the proposed model is used the TORGO speech signal database. With an accuracy score of 93.87%, CNN shows promising results in early dysarthric speech diagnosis.

Keywords

• Dysarthia
• Speech detection
• CNN
• MFCC
• Feature extraction