Automatic Speech Recognition System to Analyze Autism Spectrum Disorder in Young Children

Abstract

It's possible to learn things about a person just by listening to their voice. When trying to construct an abstract concept of a speaker, it is essential to extract significant features from audio signals that are modulation-insensitive. This research assessed how individuals with autism spectrum disorder (ASD) recognize and recall voice identity. Autism spectrum disorder is the abbreviation for autism spectrum disorder. Both the ASD group and the control group performed equally well in a task in which they were asked to choose the name of a newly-learned speaker based on his or her voice. However, the ASD group outperformed the control group in a subsequent familiarity test in which they were asked to differentiate between previously trained voices and untrained voices. Persons with ASD classified voices numerically according to the exact acoustic characteristics, whereas non - autistic individuals classified voices qualitatively depending on the acoustic patterns associated to the speakers' physical and psychological traits. Child vocalizations show potential as an objective marker of developmental problems like Autism. In typical detection systems, hand-crafted acoustic features are input into a discriminative classifier, but its accuracy and resilience are limited by the number of its training data. This research addresses using CNN-learned feature representations to classify children's speech with developmental problems. On the Child Pathological and Emotional Speech database, we compare several acoustic feature sets. CNN-based approaches perform comparably to conventional paradigms in terms of unweighted average recall.