The contribution of multiple phonetic cues to the ternary laryngeal stop contrast in Seoul and Yanbian Korean: Insights from classification in machine learning

Abstract

Previous studies have extensively explored the phonetic cues associated with the Korean ternary laryngeal stop con-trast, identifying systematic inter-series differences in voice onset time (VOT), f0, and voice quality. However, few have attempted to integrate multiple cues within a unified framework to quantify their relative contributions or to compare these patterns across dialects. This study addresses this gap by applying two supervised machine learning classification models– linear discriminant analysis (LDA) and random forest (RF)–to evaluate the contribution of VOT, f0, and glottal tension in Seoul Korean (non-tonal) and Yanbian Korean (tonal). An unsupervised K-means clustering model was also employed to assess the robustness of the results. These findings reveal distinct cue-weighting patterns be-tween the two dialects: in Seoul, f0 (primary)>VOT (primary)>glottal tension (negligible); in Yanbian, VOT (primary)>f0 (secondary)>glottal tension (secondary). These differences are interpreted in line of the absence versus presence of lexical pitch accent, with f0 being more robustly utilized in the non-tonal Seoul Korean. Beyond offering a more fine-grained account of the phonetic realization of the Korean stop triplet, this study proposes a replicable and extensible methodological framework for further research–one that may be adapted to other contrastive systems involving multiple cues.