Application of selective feature integration in acoustic feature analysis of children with autism

Abstract

In the early diagnosis of children with autism, the current main methods are based on doctor's clinical observation, eeg and ocular EEG methods, which require complex equipment and difficult acquisition process. Based on support vector machine, CNN, LSTM and other classifiers, this paper selects six acoustic features, including CQCC, PLP, LPCC, FBANK, PNCC and MFCC, for comparison and ranking according to accuracy. The traditional direct fusion method is improved by selective integration method. 2–3 features with the highest accuracy after sorting were selected for fusion again, and the accuracy and AUC of the new features were improved, reaching 99.1% in the best case. The area of AUC is close to 1, and the classification effect is good.