Multidimensional Acoustic-Prosodic Quantification Framework Using Unscripted Speech for Autism Spectrum Disorder...

Area of Science:

• Neuroscience
• Speech Science
• Developmental Psychology

Background:

• Early detection of autism spectrum disorder (ASD) is crucial for intervention, but current speech-based methods using scripted tasks are limited.
• Automatic detection of speech abnormalities in children with ASD faces challenges due to task limitations and generalizability issues.

Purpose of the Study:

• To develop and evaluate an unscripted speech-based framework for quantifying atypical acoustic-prosodic patterns for automatic ASD identification.
• To enable large-scale, non-clinical screening of ASD using naturalistic interactions.

Main Methods:

• Collected spontaneous speech from children with ASD and typically developing (TD) children during naturalistic interactions.
• Extracted multidimensional acoustic features from free-flowing conversations.
• Utilized linear discriminant analysis to model and classify ASD-related prosodic patterns.

Main Results:

• Identified significant atypical prosodic patterns in children with ASD, including reduced speech continuity and increased pitch variability.
• Achieved robust classification performance (accuracy = 0.85 ± 0.07, F1 = 0.86 ± 0.07).
• Found that atypical patterns were more pronounced in open-ended dialogues and correlated with language ability.

Conclusions:

• Unscripted speech analysis is vital for capturing atypical prosodic patterns in ASD.
• The developed framework shows promise for large-scale ASD screening outside clinical settings.
• Acoustic-prosodic patterns can serve as objective markers for assessing ASD severity.