TL;DR
This paper quantifies pitch variations in the speech of children with Autism Spectrum Disorder (ASD) and Specific Language Impairment (SLI) compared to typically developing children. They found significant differences in pitch characteristics that can be used for automatic classification, highlighting the crucial impact of pitch extraction methods on research findings.
What Problem Does This Paper Address?
Understanding the nuances of speech prosody, particularly pitch (intonation), in children with neurodevelopmental disorders like ASD is challenging. Previous research has yielded conflicting results, making it difficult to establish reliable diagnostic markers. This paper aims to address this by:
- Quantifying Pitch Differences: Precisely measuring pitch characteristics in children with ASD, SLI, and typical development (TD) to identify statistically significant differences.
- Developing Automatic Screening: Exploring whether these identified pitch features can be used to automatically classify these diagnostic groups.
- Investigating Methodological Impact: Highlighting how different choices in pitch extraction algorithms and parameters can substantially influence research outcomes, potentially explaining past discrepancies.
The heterogeneity within the ASD population is a major obstacle, so the authors also leverage a dataset that distinguishes between ASD with and without language impairment (ALN vs. ALI) to better isolate the effects of autism itself versus co-occurring language issues.
Methods
The researchers employed a quantitative analysis of speech pitch (fundamental frequency, F0). Key methodological aspects include:
- Pitch Extraction: Using the Snack toolkit's ESPS method with specific parameters: a frame shift of 10 ms, a frame size of 7.5 ms, and an allowed pitch range of 100 to 600 Hz. They emphasize the importance of reporting these settings for reproducibility.
- Feature Extraction: Two main approaches were used:
- Global Pitch Features: Calculating statistical properties from all available pitch data for a subject, essentially characterizing the per-child pitch histogram. This includes measures of location (mean, median), spread (standard deviation, median absolute deviation, interquartile range), asymmetry (skewness), and peakedness (kurtosis).
- Per-Utterance Pitch Features: Quantifying the distribution of pitch properties on a per-utterance basis.
- Statistical Analysis: A multi-pronged approach was used to identify significant group differences:
- Two-sample t-tests: Applied to individual features.
- FDR (False Discovery Rate) correction: To control for multiple comparisons and reduce spurious findings.
- 50-50 MANOVA: A robust multivariate analysis method that handles correlated features well.
- Monte Carlo simulations: Used to assess the significance of differences by comparing actual group memberships to randomly permuted ones.
- Classification: A Naïve Bayes classifier was used with a leave-one-out cross-validation scheme to evaluate the discriminative power of the identified features.
Datasets
The primary dataset used is the CSLU Autism Speech Corpus. This is a large, well-characterized dataset containing audio recordings of children aged 4-9. For this study, a subset of the corpus was analyzed, specifically focusing on speech elicited during the Autism Diagnostic Observation Schedule (ADOS) diagnostic instrument. The corpus includes children with:
- Autism Spectrum Disorder (ASD)
- Specific Language Impairment (SLI)
- Typically Developing (TD) children
- ASD with language impairment (ALI)
- ASD with normal language (ALN)
The authors also mention replicating measurements from previous studies ([8] and [10]) using their own data.
Key Insights and Ideas
- Pitch as a Biomarker: Pitch characteristics are significantly different between children with ASD/SLI and TD children, suggesting pitch could serve as a valuable, objective marker for these neurodevelopmental disorders.
- Methodological Sensitivity: The choice of pitch tracker, its parameters (like pitch ceiling), and error correction methods can dramatically alter research findings. This explains why previous studies on prosody in autism have sometimes conflicted.
- Robust vs. Non-Robust Measures: Robust statistical measures (like median and MAD) often reveal significant differences between groups even when less robust measures (like mean and SD) do not, highlighting their utility in noisy or variable data.
- ASD vs. SLI Nuances: While significant differences were found between ASD/ALN groups and TD, the study did not find significant pitch differences between the ASD subgroups (ALN vs. ALI) or between SLI and ALI. This suggests that some atypical pitch features might be common across different neurodevelopmental profiles or related to co-occurring conditions rather than specific to autism alone.
- Classification Potential: Even with a simple classifier and a limited set of features, automatic classification accuracy significantly above chance level is achievable, indicating the practical utility of these pitch features.
Key Results
- Significant Group Differences:
- ALN (ASD, normal language) vs. TD: All four statistical features related to the first moment (location) and second moment (spread) differed significantly. Skewness and kurtosis also showed significant differences.
- SLI vs. TD: Standard deviation (SD) and median absolute deviation (MAD) were significantly different, as was the median, though to a lesser extent.
- Classification Performance:
- ALN-TD: Achieved 74% accuracy, 0.86 recall, and 0.75 AUC of ROC.
- SLI-TD: Achieved 69% accuracy, 0.50 recall, and 0.72 AUC of ROC.
- These results are noted as being significantly better than chance.
Limitations and Future Work
- Methodological Reproducibility: The authors strongly emphasize the need for detailed reporting of pitch extraction methods and settings to ensure reproducibility of findings.
- Subgroup Differences: The lack of significant pitch differences between ASD subgroups (ALN vs. ALI) and between SLI and ALI suggests that further research is needed to understand the specific contributions of autism versus language impairment to pitch characteristics. It also implies that some observed atypicalities might be shared across different disorders or subtypes.
- Future Directions:
- Relating pitch features to available neurocognitive measures.
- Incorporating other prosodic aspects like loudness and rhythm.
- Identifying specific utterance curve properties that distinguish diagnostic groups.
- Investigating the relationship between per-utterance statistical features and utterance curve forms.
- Exploring more sophisticated machine learning algorithms and feature selection for improved classification.
Summary updated