New Proposed Solution for Speech Recognition without Labeled Data: Tutoring System for Children with Autism Spectrum Disorder

Abstract

Children diagnosed with autism spectrum disorder (ASD) face challenges in understanding situations, verbal communication, and social interactions. Autism can manifest differently in each child, and it can be characterized by various degrees of severity. Some common behaviors observed in children with ASD include poor skills, repetitive behaviors, delayed speech, reasoning difficulties, narrow interests, and challenges with social interactions and communication, such as recognizing social cues. As every child with ASD has unique educational needs, there is no universal solution for treating the condition. This paper aims to introduce an adaptive educational system that will help children with ASD acquire new skills and improve their communication abilities, enabling them to better integrate into society. The proposed system will be based on therapist-researched and -analyzed activities that utilize speech recognition technology. To address the resource requirements of labeled datasets, we propose a new approach that leverages generative adversarial networks (MelGAN) to produce responses that closely resemble a child's voice. This allows for the comparison of the generated response with the correct answer using similarity metrics. The system was tested on Algerian children with ASD who speak Algerian dialect, and the results were promising and this can open a new direction for developing educational systems that do not rely on labeled datasets.