From acoustic features to clinical meaning: Emerging technologies for interpretable AI in pediatric speech
TL;DR
Recent work on computer-assisted syllable analysis of continuous speech, published in the 2024 JASA special issue on Acoustic Cue-Based Perception and Production of Speech, demonstrates how linguistically structured, landmark-based acoustic representations improve robustness and interpretability relative to conventional spectral features and clinical metrics.
Abstract
Interpretable artificial intelligence (AI) has become an increasingly important topic in speech communication as data-driven methods are used to support clinical decision-making in speech health. This talk highlights acoustic technologies that advance interpretable AI for pediatric preschool-age speech. Recent work on computer-assisted syllable analysis of continuous speech, published in the 2024 JASA special issue on Acoustic Cue-Based Perception and Production of Speech (Speights et al., 2024), demonstrates how linguistically structured, landmark-based acoustic representations improve robustness and interpretability relative to conventional spectral features and clinical metrics. Advances in automatic speech recognition further enable intelligibility to be characterized using Speech Intelligibility Probabilities (SIPs), which rely on model-intrinsic uncertainty rather than transcript-based accuracy metrics. Phoneme- and language-level recognition analyses additionally reveal age-stratified confusion structure associated with speech development. Together, these developments point toward future directions in which interpretable, developmentally grounded representations play a central role in advancing speech acoustics and clinical AI for pediatric speech.