Using general sound descriptors for early autism detection

Abstract

Early detection of autism is crucial for successfully dealing with it and reduce/eliminate its effects. In other words, early treatment can make a big difference in the lives of many children with this disorder. Consequently, in this study the pattern recognition algorithms are used to determine the unique features of the voice of autistic children to distinguish between the autistic children and normal children between ages 2 and 3. These descriptors extract various audio features such as temporal features, energy features, harmonic features, perceptual and spectral features. Two feature selection methods are used and the results are compared. One method is based on comparing the effect of using all of a group features together and another method compares the effect of using features one by one. The selected features are used to classify selected children into autistic and non-autistic ones. The results show 96.17 percent accuracy. After feature selection, we classified data using S.V.M classifier for recognizing two types of input data.