Voice acoustics have been extensively investigated as potential non-invasive markers for Autism Spectrum Disorder (ASD). Although many studies report high accuracies, they typically rely on highly controlled clinical protocols that reduce linguistic variability. Their data is also recorded using specialized microphone arrays that ensure high quality recordings. Such dependencies limit their applicability in real-world or in-home screening contexts. In this work, we explore an alternative approach designed to reflect the require-ments of mobile-based applications that could assist parents in monitoring their children. We use an open-access dataset of naturalistic storytelling, extracting only the speech seg-ments in which the child is speaking. We applied previously published ASD voice-analysis pipelines to this dataset which yielded suboptimal performance under these less controlled conditions. We then introduce a deep learning–based method that learns discriminative representations directly from raw audio, eliminating the need for manual feature extraction while being more robust to environment noise. This approach achieves an accuracy of up to 77% in classifying children with ASD, children with Atten-tion Deficit Hyperactivity Disorder (ADHD), and neurotypical children. Frequency-band occlusion sensitivity analysis on the deep model revealed that ASD speech relied more heavily on the 2000–4000 Hz range, TD speech on both low (100–300 Hz) and high (4000–8000 Hz) bands, and ADHD speech on mid-frequency regions. These spectral patterns may help bring us closer to developing practical, accessible pre-screening tools for parents.