Acoustic Features Characterization of Autism Speech for Automated Detection and Classification

Abstract

The verbal children affected with autism spectrum disorder (ASD) often shows some notable acoustic patterns. This paper represents the classification of autism speech, i.e., the speech signal of children affected with ASD. In addition, this work specifically aims to classify the speech signals of non-native Indo English speakers (children) affected with ASD. Previous studies, however, have focused only on native English speakers. Hence, for this study purpose a speech signal dataset of ASD children and a speech signal dataset of normal children were recorded in English, and all the children selected for the data collection were non-native Indo English speakers. Here, for the ASD and the normal children, the acoustic features explored for classification are namely, fundamental frequency (FO), strength of excitation (SoE), formants frequencies (F1 to F5), dominant frequencies (FD1, FD2), signal energy (E), zero-crossing rate (ZCR), mel-frequency cepstral coefficients (MFCC), and linear prediction cepstrum coefficients (LPCC). Further, these feature sets are classified by utilizing different classifiers. The KNN classifier model achieves the highest 96.5% accuracy with respect to other baseline models explored here.