Raw-Speech DirichNet Model for Early Detection of Autism Spectrum Disorder in Children

Autism Spectrum Disorder (ASD) significantly af-fects communication, social interactions, and behavior. Early diagnosis in children is critical for initiating timely interventions to improve developmental outcomes. This study explores a deep learning-based method for ASD detection using raw audio signals, which contain subtle speech characteristics indicative of ASD. To address the limitations of conventional convolutional neural networks (CNNs) in capturing complex auditory features, an enhanced CNN architecture with sinc and Dirichlet layers is proposed. Sinc extracts frequency-specific features, while Dirichlet layers offer fine-tuned frequency control. Experiments on the CASD-SC dataset across Mel, Bark, and ERB scales revealed that the Dirichlet layer with N - Even configuration achieved a superior accuracy of 85.9% on the Mel scale. With performance similar to the SincNet model, the DirichNet model introduces a new approach for advancing research in ASD detection.