Summary: Researchers used machine learning to detect cross-linguistic speech patterns in children on the autism spectrum, identifying prosodic features—especially rhythm—that appear consistent across English and Cantonese.
Source: Northwestern University
Researchers at Northwestern University led a study that applied machine learning to identify speech characteristics associated with autism spectrum disorder (ASD) that are shared between English and Cantonese speakers. The findings suggest that certain prosodic features of speech could become objective markers to aid diagnosis and guide interventions.
Conducted in collaboration with a team in Hong Kong, the research offers new ways to separate genetic influences from language-specific or environmental factors that shape communication in people with autism. By focusing on measurable acoustic features rather than subjective impressions, the study provides a more consistent approach to characterizing speech differences linked to ASD.
Clinicians and researchers have long observed that many children with autism speak more slowly and show distinctive pitch, intonation and rhythm patterns compared with typically developing peers. These prosodic differences have been difficult to quantify consistently across languages, and their origins—whether genetic, neurobiological or shaped by language environment—have remained unclear.
To address this, the research team led by Molly Losh and Joseph C.Y. Lau, with collaborator Patrick Wong in Hong Kong, used supervised machine learning to analyze narrative speech samples from English- and Cantonese-speaking young people with and without autism. Participants told their versions of the story from the wordless children’s picture book “Frog, Where Are You?” and researchers extracted acoustic features relevant to rhythm and intonation from those recordings.
Published in PLOS One on June 8, 2022, the study showed that rhythm-related acoustic features were effective at classifying ASD diagnosis both within individual languages and across English and Cantonese. Intonation-related features were predictive in English but did not generalize in the same way to Cantonese, highlighting how some prosodic differences may depend on language structure.
“When you see similar speech patterns linked to autism across languages that are structurally very different, those shared traits are more likely tied to genetic liability for autism,” said Losh, the Jo Ann G. and Peter F. Dolle Professor of Learning Disabilities at Northwestern. She also noted that the variability discovered across languages could point to prosodic features that are more flexible and therefore promising targets for therapy.
Lau emphasized the role of machine learning in moving past English-language bias and human subjectivity in autism research. “This approach allowed us to identify speech elements that can predict an autism diagnosis,” he said. “Rhythm emerged as the most prominent of those features. We hope this work lays the groundwork for future machine learning–based studies of autism.”
The study points to practical applications for artificial intelligence in autism assessment and care. Automated analysis of speech could reduce the diagnostic burden on clinicians, increase accessibility of evaluation in diverse communities, and offer objective measures that are comparable across languages. Because computers quantify acoustic characteristics rather than relying on subjective impressions, such tools could support cross-cultural research and clinical practice.
Because some features identified by the machine-learning models were shared across English and Cantonese while others were language-specific, the authors believe the approach could help design targeted interventions. Machine learning could identify which speech features respond to therapy and then track a speaker’s progress objectively over time.
Beyond clinical applications, the findings may aid efforts to pinpoint genetic and neural mechanisms linked to speech processing differences in autism. “One brain network that is involved is the auditory pathway at the subcortical level, which is robustly tied to differences in how speech sounds are processed by individuals with autism compared with typically developing individuals across cultures,” Lau said. The team aims to explore whether those early processing differences are responsible for the behavioral speech patterns observed and to investigate their neural genetics.
About this AI and ASD research news
Author: Max Witynski
Source: Northwestern University
Contact: Max Witynski – Northwestern University
Image: The image is in the public domain
Original Research: Open access. “Cross-linguistic patterns of speech prosodic differences in autism: A machine learning study” by Joseph C. Y. Lau et al. PLOS ONE
Abstract
Cross-linguistic patterns of speech prosodic differences in autism: A machine learning study
Differences in speech prosody are a commonly observed feature of Autism Spectrum Disorder (ASD). Yet it has been unclear how these prosodic differences present across languages that differ substantially in their prosodic systems.
Using a supervised machine-learning approach, the researchers examined acoustic features related to rhythmic and intonational aspects of prosody, derived from narrative samples in English and Cantonese—two typologically and prosodically distinct languages.
The models achieved reliable classification of ASD diagnosis using rhythm-related features both within each language and across languages. Intonation-related features supported classification in English but did not generalize to Cantonese in the same way.
These results highlight rhythm as a core prosodic feature affected in ASD while also demonstrating that other prosodic properties, such as intonation, may vary depending on language-specific characteristics. The findings point toward objective, language-aware methods for studying speech differences in autism and for developing diagnostic and therapeutic tools.