Identifying stable speech-language markers of autism in children: Preliminary evidence from a longitudinal telephony-based study

Sunghye Cho, Riccardo Fusaroli, Maggie Rose Pelella, Kimberly Tena, Azia Knox, Aili Hauptmann, Maxine Covello, Alison Russell, Judith Miller, Alison Hulink, Jennifer Uzokwe, Kevin Walker, James Fiumara, Juhi Pandey, Christopher Chatham, Christopher Cieri, Robert Schultz, Mark Liberman, Julia Parish-morrisNAACL 2022

TL;DR

This study explores the use of telephony to collect speech data from autistic and neurotypical (NT) children, identifying stable acoustic and lexical differences that could serve as early markers for autism. The findings suggest that specific vocal characteristics and language patterns are consistently present in autistic children, even when data is collected remotely, paving the way for scalable intervention monitoring.

What Problem Does This Paper Address?

Diagnosing and monitoring autism spectrum disorder (ASD) in children often relies on in-person assessments, which can be costly, time-consuming, and inaccessible for many families, particularly those in geographically isolated or underserved communities. Furthermore, the presence of unfamiliar adults during assessments can influence the behavior and speech of autistic children, potentially affecting the reliability of the data. There's a need for accessible, scalable, and reliable methods to identify and track developmental changes and intervention responses in autistic children using speech as a primary indicator. This paper addresses this by investigating whether speech and language features, collected remotely via a telephony system, can reliably distinguish between autistic and neurotypical children and remain stable over time.

Methods

The researchers developed a telephony platform to collect speech samples from children. This platform utilized a high-availability server, VoIP services, and telephony software to support both single and dual speaker modes. For this study, they focused on a "picture description task" administered over two weekly sessions.

Key methodological aspects include:

  • Longitudinal Data Collection: Weekly speech samples were collected over two sessions to assess the stability of speech features over time.
  • Telephony-Based Elicitation: A novel approach using a phone-based system to improve accessibility and reduce the burden of in-person visits.
  • Semi-Structured Task: Children described different pictures in each session, providing a consistent yet varied stimulus for speech production.
  • Data Annotation and Feature Extraction: Recordings were transcribed, and then analyzed for both acoustic and textual features.
  • Acoustic Features: Low-level descriptors like pitch, jitter (frequency variability), shimmer (amplitude variability), harmonic-to-noise ratio (HNR), and spectral moments (centroid, standard deviation, skewness, kurtosis) were extracted. Pitch was normalized to semitones using individual 10th percentiles to account for physiological differences.
  • Textual Features: Words were tagged for part-of-speech (POS) categories. Counts of POS categories, fillers, partial words, repetitions, and "hm" were normalized per 100 words. Content words were also analyzed for age of acquisition, concreteness, ambiguity, and familiarity. LIWC (Linguistic Inquiry and Word Count) categories were also analyzed.
  • Statistical Analysis: Analysis of covariance (ANCOVA) models were used to compare groups, with session and group-session interaction as independent variables, and sex as a covariate.

Datasets

The study utilized data collected from 29 children (13 autistic, 16 neurotypical) who completed two sessions of the telephony-based speech collection. The participants were matched on age, full-scale IQ, and self-reported race. The paper does not introduce a new benchmark dataset but rather uses data collected through their novel telephony protocol for analysis.

Key Insights and Ideas

  • Telephony as a Viable Data Collection Tool: The study demonstrates that a telephony system can yield high-quality recordings suitable for detailed acoustic and linguistic analysis, offering a scalable solution for remote data collection.
  • Stable Group Differences: Crucially, the study found stable group differences in certain acoustic and lexical features between autistic and NT children across two sessions. This stability is vital for identifying reliable markers.
  • Acoustic Markers of Autism: Autistic children exhibited higher median jitter and shimmer, and greater spectral skewness and kurtosis. These vocal characteristics, related to voice quality and timber, appear to be consistent markers.
  • Durational and Lexical Differences: Autistic children showed longer speech segments and shorter, less frequent pauses. Lexically, they produced fewer conjunctions and pronouns and used content words with higher age of acquisition.
  • Session Effects Highlight Context: Many lexical features differed significantly between sessions, suggesting that the specific pictures used in the task had a substantial influence on word choice. This underscores the importance of careful stimulus design in future research.

Key Results

  • Acoustic: Autistic children had significantly higher median jitter (p=0.052) and shimmer (p=0.046). They also showed significantly higher mean and median spectral skewness (p<0.001 for mean, p=0.016 for median) and kurtosis (p<0.001 for mean, p=0.035 for median). Pitch and HNR did not differ significantly between groups.
  • Durational: Autistic children produced longer (p=0.007) and more variable (p=0.005) speech segment durations. Total pause duration was shorter in autistic participants (p=0.028), and they paused less frequently (p=0.015).
  • Lexical: Autistic participants produced fewer conjunctions (p=0.029) and pronouns (p=0.034) than NT children. Their content words had a higher age of acquisition (p=0.015). They also used fewer perception-related words (p=0.004) and more time-related words (p=0.033).

Limitations and Future Work

  • Sample Size: The study is described as a "preliminary report" with a relatively small sample size (n=29). The authors plan to study these features further in a larger sample.
  • Stimulus Influence: The significant session-based differences in lexical features highlight that picture selection can heavily influence word-level analysis. Future research needs to carefully design stimuli to minimize this effect, potentially focusing on more abstract lexical features like pronoun use.
  • Sex Ratio: The groups were not matched on sex, which is noted as expected due to the higher prevalence of ASD in boys. The authors are addressing this with targeted recruitment.
  • Future Directions: The authors aim to investigate whether group differences in more abstract lexical features remain stable over all seven sessions. They also plan to explore whether the identified acoustic features could serve as validated speech markers of autism and be useful for intervention research. The potential of telephony for monitoring intervention responses and longitudinal developmental changes is a key future focus.

Summary updated

BibTeX
@inproceedings{cho2022identifying,
  author = {Sunghye Cho and Riccardo Fusaroli and Maggie Rose Pelella and Kimberly Tena and Azia Knox and Aili Hauptmann and Maxine Covello and Alison Russell and Judith Miller and Alison Hulink and Jennifer Uzokwe and Kevin Walker and James Fiumara and Juhi Pandey and Christopher Chatham and Christopher Cieri and Robert Schultz and Mark Liberman and Julia Parish-morris},
  title = {{Identifying stable speech-language markers of autism in children: Preliminary evidence from a longitudinal telephony-based study}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
}
Figure