Machine Learning Classification of Autistic Adults Using Vowel Acoustic Features

Autism is a neurodevelopmental condition for which timely and accurate identification remains an important clinical priority. Early and reliable identification can facilitate access to assessment, diagnosis, and appropriate support; however, current diagnostic pathways still rely largely on behavioural evaluation and clinical judgement. In this context, machine learning (ML) approaches have attracted growing interest because they can identify subtle and complex patterns in speech data that may not be readily captured through conventional methods. The current study investigated the potential of ML models to distinguish vowel productions from autistic and neurotypical adults based on acoustic speech features. Acoustic measures included fundamental frequency (F0), the first three formants (F1, F2, F3), duration, jitter, shimmer, harmonics-to-noise ratio (HNR), and intensity, elicited through a controlled production task. Four supervised ML models were evaluated: LightGBM, Random Forest, Support Vector Machine, and XGBoost. Under random token-level splitting, all models demonstrated good classification performance, with the best-performing model achieving an area under the curve (AUC) of approximately 0.89. However, performance decreased substantially under speaker-independent cross-validation, with AUCs approximately 0.60 and wider confidence intervals, indicating more limited generalization to vowel productions from unseen speakers. SHAP analyses nevertheless showed a broadly consistent feature-importance pattern across validation schemes, with F0 emerging as the strongest predictor, followed by intensity, while F3 and other acoustic measures made smaller contributions. These findings indicate that vowel acoustics contain information relevant to distinguishing autistic and neurotypical speech within a controlled dataset, while also demonstrating that classification performance is substantially attenuated under speaker-independent validation. The results highlight both the potential of interpretable speech-based ML for investigating acoustic markers of autism and the importance of rigorous speaker-independent evaluation.