Computational Modeling of Error Patterns in Children Speech

Abstract

This study aims to analyze the phonological processes in the speech of Dutch children, focusing on both typically developing (TD) children and those with developmental language disorder (DLD). Utilizing a dataset from a non-word repetition task, we investigate some phonological errors made by children aged 3;0 to 6;2 (years; months). Our approach involves using a combination of Levenshtein Distance and Breadth-First Search algorithms to quantify and document four common phonological processes (ie. error patterns): final consonant deletion, stopping, fronting, and gliding.

We perform statistical analyses to compare the frequency of these processes between TD and DLD children and to assess differences in the percentage of pseudowords presented and repeated by each group. Building on this analysis, we apply Optimality Theory constraints and train a maximum entropy (MaxEnt) model to evaluate each child’s pronunciation. This model is trained on TD children’s data and tested on both TD and DLD children to determine the probability of typical pronunciation patterns.The effectiveness of the classifier is assessed using receiver operating characteristic curves to distinguish between TD and DLD children.

Our findings indicate significant differences in the phonological processes use and number of words repeated (either correctly or incorrectly) between the two groups, supporting the utility of these metrics in diagnosing DLD. Additionally, the MaxEnt model demonstrates high reliability especially in the oldest group of children.

This research contributes to clinical practice by offering detailed analyses of child pronunciation errors and improving diagnostic accuracy for DLD. Theoretically, it advances our understanding of phonological development and the applicability of OT in language acquisition studies.