Self-Supervised Speech Representations in a Pre-train Speech Model Represent Key Rapid Automatized Naming Variability in Autism
University of Massachusetts (UMass) Amherst · 14.06.2025
DOI: 10.7275/scil.3173 ↗Kurz gefasst
Fachartikel in der Zeitschrift University of Massachusetts (UMass) Amherst, erschienen am 14.06.2025. Verfasst von Sarah B. Ethridge und 6 weiteren Autor:innen.
Abstract (Originalsprache)
Individuals with autism experience significant yet nuanced difficulties with pragmatic language, such as differences in speech prosody (e.g., rate, rhythm, intonation) and speech-gaze coordination, which can be challenging to measure quantitatively without powerful tools. Thus, fine-grained, comprehensive characterization of speech and related cognitive-linguistic skills is important for informing intervention strategies grounded in a clearer etiological understanding of pragmatic differences in autism. This study used Hidden-unit Bidirectional Encoder Representations from Transformers (HuBERT), a state-of-the-art, self-supervised, pre-trained speech model, to represent speech and gaze differences during a rapid naming task (RAN) in autistic individuals relative to non-autistic controls. Using Pearson’s correlations, average distance metrics were analyzed for associations with acoustic (e.g., speech rate), performance-based (e.g., naming time), and gaze metrics (e.g., regressions) of RAN, to examine the potential link between HuBERT distance measures and the attentional coordination of speech and gaze. Analyses revealed that the HuBERT distance metric was significantly correlated with more speech errors, slower speech rate, longer naming time, and more visual regressions. Together, findings demonstrate the utility of self-supervised pre-trained speech models, such as HuBERT, to capture meaningful variability in the cognitive-linguistic patterns of autism without need for pre-defined acoustic features or speech-to-text alignment.
Schlagwörter
Zitieren
APA
Ethridge, S. B., Lau, J., Chernyak, B. R., Voigt, R., Goldrick, M., Keshet, J., & Losh, M. C. (2025). Self-Supervised Speech Representations in a Pre-train Speech Model Represent Key Rapid Automatized Naming Variability in Autism. University of Massachusetts (UMass) Amherst. https://doi.org/10.7275/scil.3173
Vancouver
Ethridge SB, Lau J, Chernyak BR, Voigt R, Goldrick M, Keshet J, et al. Self-Supervised Speech Representations in a Pre-train Speech Model Represent Key Rapid Automatized Naming Variability in Autism. University of Massachusetts (UMass) Amherst. 2025. doi:10.7275/scil.3173
BibTeX
@article{Ethridge2025,
author = {Ethridge, Sarah B. and Lau, Joe and Chernyak, Bronya Roni and Voigt, Robert and Goldrick, Matt and Keshet, Joseph and Losh, Molly C.},
title = {Self-Supervised Speech Representations in a Pre-train Speech Model Represent Key Rapid Automatized Naming Variability in Autism},
journal = {University of Massachusetts (UMass) Amherst},
year = {2025},
doi = {10.7275/scil.3173}
}
EUVIMED