Speech Pattern Disorders in Verbally Fluent Individuals with Autism Spectrum Disorder: A Machine Learning Analysis

Chuanbo Hu^1*Jacob Thrasher^2*Wenqi Li¹Mindi Ruan¹Xiangxu Yu³Lynn K. Paul⁴Shuo Wang³Xin Li^1*

• ¹University at Albany, Albany, United States
• ²West Virginia University, Morgantown, United States
• ³Washington University in St Louis School of Medicine Mallinckrodt Institute of Radiology, St. Louis, United States
• ⁴California Institute of Technology, Pasadena, United States

Diagnosing Autism Spectrum Disorder (ASD) in verbally fluent individuals, based on speech patterns in examiner-patient dialogues presents significant challenges due to the subtle and varied nature of speech-related symptoms. This study analyzes recorded dialogues using the Autism Diagnostic Observation Schedule (ADOS-2) to identify distinctive speech characteristics. We extracted 40 speech-related features, categorized into intonation, volume, rate, pauses, spectral characteristics, chroma, and duration. These features, analyzed using advanced speech tools, captured complex speech patterns associated with ASD. We then trained machine-learning models to classify ASD participants into two subgroups—those with A2-defined speech pattern abnormalities and those without—and achieved an F1-score of 84.49%. We removed Mel-Frequency Cepstral Coefficients (MFCC) and Chroma features to focus on prosodic, rhythmic, energy, and selected spectral features aligned with the ADOS-2 Module 4 A2 score (speech abnormalities), aiming to reduce redundancy and balance feature contributions. This reduced feature set improved performance, with an accuracy of 85.77% and an F1-score of 86.27%, highlighting the effectiveness of a diverse combination of non-spectral and selected spectral features for characterizing speech abnormalities in ASD. While spectral features (e.g., spectral spread and spectral centroid) emerged as key features in the reduced feature set. In addition, MFCC 6 and Chroma 4 also significantly contributed to classification performance in the full feature set, indicating their role in capturing fine-grained speech variations. Together, these findings support developing context-aware models that aid the characterization of speech abnormalities in verbally fluent individuals with ASD and may complement existing clinical assessments.