Deep Learning based Behavioral Analysis and Exploration of Emotions in ASD Children

Voice based emotional analysis has become one of the most challenging and fastest advancing methodologies in the field of research. This research mainly focus on detecting and predicting the traits of Autism Spectrum Disorder (ASD) in young children by using prime cepstral features namely Mel Frequency Cepstral Coefficients (MFCC) and Mel Spectrograms (MS) which are extracted from their voices. The proposed 2-D CNN-EAA (Convolutional Neural Network for Emotion Analysis in Autism) model uses these cepstral features to categorize the eight basic emotions and recognize their variations in ASD and Typically Developing (TD) children. This analysis focuses on finding the predominant emotions and the next probable emotion in both the groups of children. The system examines the flow of emotions in both groups of children to find the disparities in their behavior and understand their perception of various situations. The proposed model is compared to machine learning techniques (SVM and MLP) and found to achieve significant performance in the process of emotion analysis. Henceforth, this research aids in better understanding their nature and assists in deciding the timely training and care to help shape their lives.