From scalar confidence to structured uncertainty: Modeling child speech intelligibility from acoustic model evidence

Abstract

Accurate assessment of children’s speech intelligibility is critical for identifying speech sound disorders. Clinical practice relies on human transcription and perceptual ratings that are costly, inconsistent, and difficult to scale. Existing ASR-based intelligibility metrics offer scalability but face two fundamental issues: (i) training-data mismatch, as adult-centric models produce unstable representations for developmentally variable child speech, and (ii) decoding artifacts, e.g., softmax sharpening, that yield spuriously confident scores and mask acoustic ambiguity. We propose a Speech Intelligibility Probability (SIP) framework that conceptualizes intelligibility as acoustically induced model uncertainty rather than as word error or decoder confidence. Using the Speech Exemplar and Evaluation Database, we extracted token-level posterior trajectories from Whisper Large-v2 under controlled decoding and analyzed them using a multidimensional uncertainty profile capturing distributional dispersion (Tsallis entropy) and temporal instability (coherence), thereby characterizing acoustic-driven uncertainty beyond post-decoding confidence scores tied to committed hypotheses. We compared SIP against Hustad et al. (2021) developmental growth curves. It aligned with human perceptual difficulty: typically developing speech clustered near mid-range expectations, while disordered speech concentrated in the bottom decile. This correspondence indicates that internal model uncertainty serves as a robust proxy for speech intelligibility.