Abstract
Atypical prosody—the unusual rhythm and melody of speech—is a hallmark of Autism Spectrum Disorder (ASD), yet it remains difficult to quantify objectively. This paper presents a computational approach to measuring these patterns by analyzing acoustic features and turn-taking dynamics in clinical interactions. Our findings reveal that children with higher ASD severity exhibit increased prosodic variability and decreased coordination between speech subsystems. Furthermore, we demonstrate that clinicians unconsciously adapt their own speech patterns in response to the child's communication style, providing a novel "interlocutor effect" metric for severity assessment.
Introduction
Clinical assessments of ASD currently rely heavily on subjective human observation to identify "atypical prosody," which can range from monotone to "sing-songy" speech. This subjectivity leads to low inter-rater reliability and lacks the precision needed for fine-grained tracking of developmental progress. This study aims to bridge this gap by developing objective, signal-derived measures that capture the multidimensional nature of speech timing, melody, and social interaction.
Methods
The researchers extracted five categories of features from audio recordings of semi-structured clinical interviews:
- Turn-taking & Speaking Rate: Metrics including response latency, speaking rate (syllables per second), and pause frequency.
- Segmental Pitch Cues: Syllable-level analysis using second-order polynomials to capture micro-prosodic idiosyncrasies in pitch slope and curvature.
- Suprasegmental Intonation: Long-term melodic movements modeled via the Momel (Modeling Melody) and Intsint algorithms. These convert pitch contours into symbolic representations such as "High," "Low," or "Same" tones.
- Goodness of Pronunciation (GOP): A measure of phonetic precision calculated using Kaldi acoustic models to determine deviation from standard pronunciation.
- Prosodic Coordination: A novel feature group measuring the correlation between pitch, duration, and intensity within a syllable, testing the hypothesis that typical speakers synchronize these elements for emphasis.
Dataset
The study utilized a sample of 176 children participating in ADOS (Autism Diagnostic Observation Schedule) Module 3 sessions. The cohort included 95 children with a confirmed ASD diagnosis and 81 control children with other neurodevelopmental disorders (e.g., ADHD, language delays). This control group provides a rigorous baseline for distinguishing ASD-specific markers from general developmental delays.
Key Insights and Results
The analysis yielded several critical findings regarding the nature of atypical prosody and its manifestation in social contexts.
The Interlocutor Effect
One of the most significant findings is that the psychologist’s behavior is a strong predictor of the child's ASD severity. When interacting with children with higher severity, psychologists exhibited increased latency, more frequent pauses, and higher pitch variability. This suggests an unconscious adaptation by the clinician to bridge communication gaps or manage uncertainty in the interaction.
Variability vs. Monotony
Contrary to the common perception of ASD speech as "monotone," the data showed that children with higher severity actually exhibited increased macro-prosodic variability. These children used more extreme "High" and "Low" symbolic tones and showed larger pitch swings compared to their peers.
Coordination and Pronunciation
The "Prosodic Coordination" feature revealed that higher ASD severity correlates with lower synchronization between pitch and duration. While typical speakers often increase both pitch and duration to emphasize a word, these modalities appear decoupled in children with ASD. Additionally, lower "Goodness of Pronunciation" scores correlated with higher severity, suggesting that motor-articulatory control is linked to the perception of atypical speech.
Performance Summary
The combined feature set achieved a 0.35 Spearman’s correlation with ADOS severity scores. In the classification task (ASD vs. non-ASD developmental disorders), the models reached an Unweighted Average Recall (UAR) of 59%. While seemingly modest, this performance is notable given the high similarity between the ASD and control groups. The most robust predictors were the child’s segmental pitch features and the psychologist’s suprasegmental intonation patterns.
Conclusion
This work demonstrates the potential for objective acoustic measures to augment clinical ASD assessments. By focusing on interpretable features rather than "black-box" models, the researchers provide a framework that clinicians can use to understand and track specific prosodic deficits. Future work will focus on correlating these features directly with human perceptions of speech "awkwardness" and improving recording robustness across varied clinical environments.
Summary updated