Abstract: Aiming at the problems of high time consuming and cost in traditional diagnosis of dysarthria speech, a computer automatic recognition method for dysarthria is proposed. Combining the Gammatone Frequency Cepstrum Coefficients (GFCC) with the common acoustic features to form a combined acoustic feature, a differential evolution algorithm is applied for feature selection, and a logistic regression classifier is used to identify the dysarthria speech. The Torgo database is divided into three subsets, which are non-words, short words, restricted sentence. 24-dimensional GFCC and 37-dimensional commonly used acoustic features are extracted to form combined acoustic features. Finally, differential evolution algorithm and logistic regression classifier are used for identificaiton of dysarthria. Experiments show that the differential evolution algorithm can effectively select feature subsets with better ability to distinguish dysarthria and healthy speech, which can significantly improve performance in the classification of dysarthria. The experiment on non-word subsets achieves 98.18% of accuracy, 98.3% of recall, and 98.3% of precision.

Key words: GFCC, differential evolution algorithm, logistic regression, dysarthria recognition