Detecting Autism Spectrum Disorder from Raw Speech in Children using STFT Layered CNN Model

Autism spectrum disorder (ASD), a lifelong neurodevelopmental condition impacting perception, communication, and behavior, serves as the focal point of this study. It encompasses a wide range of symptoms and intensities and usually manifests in early childhood due to genetic, environmental and immunological factors. The paper introduces a new dataset for children ASD speech corpus (CASD-SC), employing a customized log spectrogram short-time Fourier transform (STFT) layered convolutional neural networks (CNN) model. The investigation encompasses the evaluation of two distinct models: the traditional feature-based CNN model and the raw waveform-based CNN model. Various configurations of the CNN's first layer, such as spectrogram and log spectrogram, are investigated, and the log spectrogram raw waveform-based CNN model is found to be 86.6 % accurate in detecting ASD. This research holds significance because it addresses a gap in the existing literature by examining child-specific raw speech data analysis. This study includes a customized log spectrogram layer, this customization emphasizes processing and training efficiency by demonstrating the benefit of utilizing raw waveforms over feature extraction. This study advances in detection of ASD in children, helping in the early diagnosis and treatment of children with ASD.