A whole-word computational algorithm for phonological error scoring: Differentiating children with developmental language disorder, hearing impairment, and typical development

Accurate phonological assessment is critical for the management of phonological disorders, yet current methods rely heavily on time-intensive manual transcription and analysis. We introduce a computational phonological scoring tool for phonological production in two groups that are at a high risk for phonological disorder, namely children with developmental language disorder (DLD) and hearing impairment (HI). We test whether the computational tool supports (a) identification of phonological problems vs. typical development (TD) and (b) differentiation between DLD and HI. Seventy-two Swedish-speaking children (29 DLD, 14 HI, 29 TD) completed a picture naming task. Phoneme-level distances between targets and productions were computed with Normalised Damerau-Levenshtein Distance, yielding a Composite Phonological Score (CPS) and counts of deletions, insertions, substitutions, and transpositions. Machine-learning models evaluated classification performance. DLD and HI groups showed higher CPS (more errors) than TD. CPS did not differ between DLD and HI, but deletions were more frequent in HI than DLD. Feature importance analysis revealed that CPS was the most accurate metric for distinguishing children with impairments (DLD and HI) from TD peers, underscoring its utility as a primary screening marker. Distinguishing DLD vs. HI required the detailed error profile and broader language measures rather than CPS alone. Our computational scoring of whole-word phonological accuracy provides more information than the standard correct or incorrect scores, supporting screening for phonological problems, while group-specific error patterns (e.g. deletions) and additional language measures may aid differentiation between DLD and HI.