Author(s) :

Krishna Pabbu.

Volume/Issue :

Abstract :

Early screening for autism spectrum disorder (ASD) is still difficult. Current assessments take time, depend heavily on clinicians, and are hard to scale. Speech transcripts are easier to collect and avoid many privacy concerns. This study aims to evaluate whether simple linguistic features and Zipfian wordfrequency metrics can distinguish children with ASD from typically developing children using short speech transcripts. Prior work reports classification accuracies of 75–80% using linguistic features, and some studies suggest that ASD speech may deviate from Zipfian word-frequency patterns. This study tests those claims using the Eigsti corpus, a balanced subset from ASDBank, which includes 16 autistic participants and 16 typically developing participants. Neural networks were trained on linguistic features, Zipf-based metrics, and their combination across 100 randomized train-test splits with data augmentation. Performance remained near chance level, approximately 50% accuracy, across all conditions. Feature analysis showed that ASD participants used fewer conjunctions, while other linguistic and Zipfian features did not show consistent separation or predictive value. Overall, these results do not support earlier reports. Linguistic and Zipfian features alone are not enough for reliable ASD classification from short transcripts. The results also highlight reproducibility challenges in small datasets and suggest that larger datasets and multimodal approaches will be needed.