Automated ASD detection in children from raw speech using customized STFT-CNN model

Autism spectrum disorder (ASD), a prevalent neurodevelopmental condition impacting cognitive, communicative, and behavioral aspects, typically manifests in early childhood due to genetic, environmental, and immunological factors. Employing a novel dataset termed children’s ASD speech corpus (CASD-SC), the research makes use of short-time Fourier transform (STFT) layered convolutional neural networks (CNN), incorporating an image input layer and a sequence input layer. The analysis encompasses data both with and without augmentation, exploring various CNN configurations. Results showcase that the log spectrogram-based STFT layered CNN model achieves 86.6% accuracy for the raw data, while the pre-emphasis filter (PEF) with learnables-based STFT layered CNN model attains 99.1% accuracy for the data with augmentation for detecting ASD in children. This investigation bridges the literature gap by evaluating child-specific raw speech data. The study underscores the significance of processing and training efficiency in ASD diagnosis and promotes early intervention techniques by improving ASD detection in children.