Research on Acoustic Feature Extraction of Crying for Early Screening of Children with Autism

Abstract

In the field of early diagnosis of autism children, the current method is mainly based on doctors' clinical observation and experience with the assistance of some quantitive indexes. In this paper, we propose to use acoustic features of crying sound for the early diagnosis of autistic children. Four acoustic features extraction methods (wavelet decomposition coefficient, DWT-MFCC, MFCC and LPCC) and two machine leaning-based classifiers (SVM, CNN) are applied to the crying sounds of autism children aged 2 to 3 years old. Comparison experiments show that MFCC features with SVM and CNN achieved the highest recognition rate, while DWT-MFCC exhibited the most stable performance in the case of five different SNRs. And it also can be seen from the experiments that the convergence speed of MFCC and DWT-MFCC features with CNN is approximately the same. Our research may ultimately help doctors diagnose autism in young children from a speech signal processing perspective.