Multidimensional Acoustic–Prosodic Quantification Framework Using Unscripted Speech for Autism Spectrum Disorder Identification

Although clinical observations have noted early speech abnormalities in children with autism spectrum disorder (ASD), automatic speech-based detection remains challenging. This is primarily due to the reliance on scripted tasks, which younger children often struggle to complete and which are not generalizable to large-scale, non-clinical screening. To address this, we developed an unscripted speech-based framework to quantify atypical acoustic-prosodic patterns for automatic ASD identification in naturalistic interactions. It processes free-flowing conversations, extracts multidimensional acoustic features from the time and frequency domains, and models ASD-related prosodic patterns for classification. For evaluation, we collected spontaneous speech from 88 children with ASD (3-10 years) and 82 typically developing (TD) children (3-9 years) during naturalistic interactions on daily topics (e.g., toys, animated movies, storybook reading). Group comparisons revealed atypical prosodic patterns in ASD, including reduced speech continuity, speech rate, and Formant 3, alongside increased zero-crossing rate, pitch, pitch variability, and Formant 1 (all p < 0.01). Using these features, a linear discriminant analysis classifier achieved robust performance (accuracy = 0.85 ± 0.07, F1 = 0.86 ± 0.07). Further analyses indicated no significant gender interaction (p > 0.05), but a pronounced effect of speech context (p < 0.01), with atypical patterns being more evident in open-ended dialogues than in text-guided settings. Moreover, these patterns correlated with clinical scores (p < 0.05), particularly language ability, demonstrating the framework's utility for assessing ASD severity. These findings underscore the importance of analyzing unscripted speech to capture atypical prosodic patterns and provide a basis for large-scale ASD screening outside clinical settings.