Automated Sinhala Speech Emotions Analysis Tool for Autism Children

Abstract

— Autism Spectrum Disorder (ASD) is a neurological disorder that impairs children's development and symptoms that can be noticed in early childhood. One of the main diagnosis characteristics of ASD is the child having unusual emotions and expressions during social interactions. The main problem is how to distinguish these symptoms. Only 14 out of 100 Autistic kids, before they reach the age of 24 months, get medical treatments since the unavailability of resources to identify them early. If they can be recognized early, a therapeutic engagement can be done to help them overcome those issues in social interactions, when they reach school-going age. The focus of this research is to develop a tool to screen atypical children from typical children. This research attempts to recognize the correct emotion of a child, while the child is talking. The input audio stream of children was normalized into a specific range, sub-framed into 2s length for language-independent, noise reduction, and age independence features, and extracting the most effective 40 audio features. The Convolutional Neural Network (CNN) based model classifies eight different emotions of sad, disgust, surprise, neutral, happy, calm, fear, and angry with an accuracy matrix of F1 score of 0.90, even in the uncontrol environment. If the classifying emotions have small frequency variances, the trained model has the ability to handle them.