Towards Precision Characterization of Communication Disorders using Models of Perceived Pragmatic Similarity

Nigel G. Ward, Andres Segura, Georgina Bugarini, Heike Lehnert-LeHouillier, Dancheng Liu, Jinjun Xiong, Olac Fuentes

Abstract

Traditional diagnostic models for communication disorders often rely on categorical labels and large datasets, frequently overlooking the nuanced social and functional use of language known as pragmatics. This paper introduces a novel approach using "perceived pragmatic similarity" to characterize disorders such as Autism Spectrum Disorder (ASD) and Specific Language Impairment (SLI). By utilizing self-supervised speech models (HuBERT), we demonstrate that a general-purpose similarity metric can identify atypical speech patterns and support precision medicine, even in low-resource clinical settings.

Introduction

Speech technology for communication disorders typically faces a "categorical trap," attempting to force diverse symptoms into rigid binary labels. This approach is limited by the scarcity of labeled clinical data and the high variability of symptom manifestation. Furthermore, existing models often focus on phonetics or syntax, neglecting pragmatics—the social context of language. We propose a tool that characterizes speech based on how it is perceived socially, bridging the gap between clinical diagnosis and functional communication.

Methodology

Our core approach involves estimating pragmatic similarity between pairs of utterances.

  • Model Architecture: We utilize a modified version of a model by Segura, extracting features from the 24th layer of a pretrained HuBERT model.
  • Similarity Metric: Feature vectors for each utterance are compared using cosine similarity to determine pragmatic proximity.
  • Classification: We employ a k-Nearest Neighbor (kNN) approach. A speaker's utterances are compared to a reference set; if they are consistently closer to the ASD group than the neurotypical (NT) group, they are classified as ASD.
  • Zero-Shot Atypicality: We also evaluate a "centroid" model that measures the distance of a speaker's average vector from a "normal" population, enabling detection of atypicality without disorder-specific training.

Experimental Setup and Datasets

The model was evaluated across three datasets:

  1. DRAL Corpus: 2,893 English utterances from natural conversations used to validate the model against human pragmatic judgments.
  2. NMSU Dataset: Recordings of 28 age-matched adolescents (ASD vs. NT) performing a "find-the-difference" task.
  3. Edmonton Narrative Norms Instrument (ENNI): Audio of children (ages 4–10) with SLI and Typically Developing (TD) peers retelling stories.

Results and Discussion

The model's similarity predictions showed strong alignment with human judgments, nearly matching human inter-annotator agreement. In clinical tasks, the similarity-based approach proved highly effective.

Autism Spectrum Disorder Detection

Using the kNN approach on HuBERT features, the model achieved 82% accuracy in discriminating between ASD and neurotypical adolescents. This performance highlights the diagnostic signal present in pragmatic similarity.

Confusion matrix for ASD vs. NT classification showing 82% accuracy.

Specific Language Impairment Discrimination

The method also outperformed standard baselines (such as utterance length) in identifying SLI in children. The confusion matrix below illustrates the model's ability to distinguish SLI speakers from typically developing peers.

Confusion matrix for SLI vs. TD classification.

Precision Characterization

Beyond binary classification, the "typicality" benchmark (using a neurotypical centroid) allows clinicians to identify departures from standard speech patterns. This supports "precision characterization," where clinicians can find comparable utterances from other patients to inform personalized treatment strategies.

Limitations and Future Work

Current limitations include the model's focus on single utterances, ignoring broader conversational context. Future work will integrate non-verbal cues (gestures, facial expressions) and lexical content to develop a more holistic and explainable model of pragmatic behavior.

Conclusion

By modeling perceived pragmatic similarity, we move toward a more nuanced and data-efficient method for characterizing communication disorders. This approach avoids the pitfalls of rigid categorization and provides a foundation for precision medicine in speech-language pathology.

Summary updated

BibTeX
@inproceedings{ward2025towards,
  author = {Nigel G. Ward and Andres Segura and Georgina Bugarini and Heike Lehnert-LeHouillier and Dancheng Liu and Jinjun Xiong and Olac Fuentes},
  title = {{Towards Precision Characterization of Communication Disorders using Models of Perceived Pragmatic Similarity}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2025},
  doi = {10.1109/icassp49660.2025.10890858},
}
Figure