Classifying Speech of ASD Affected and Normal Children Using Acoustic Features

Abstract

Children affected with autism spectrum disorder (ASD) produce speech that consists of distinctive acoustic patterns, as compared to normal children. Hence, acoustic analyses can help classifying speech of ASD affected children from that of normal children. In this study, the aim is to identify those discriminating characteristics of speech production that help classification between speech of children with ASD and normal children. Two separate datasets were recorded for this study: the English speech of children affected with ASD and the English speech of normal children. Comparative analyses of acoustic features derived for both datasets are carried out. Changes in the speech production characteristics are examined in three parts. Firstly, changes in the excitation source features F0 and strength of excitation (SoE) are analyzed. Secondly, changes in the vocal tract filter features the formants (F1 to F5) and dominant frequencies (FD1, FD2) are analyzed. Thirdly, changes in the combined source-filter features signal energy and zero-crossing rate are analyzed. Different combinations of the feature sets are then classified using three different classifiers for validation of results: SVM, KNN and ensemble classifiers. Performance evaluation is carried using different combinations of features sets and classifiers. Results up to 97.1% are obtained for classification accuracy between speech of ASD affected children and normal children, using a combination of feature set with SVM classifier. The results are better than other similar few studies. This study should be helpful in developing an automated system for identffying ASD speech, in future.