Decoding the dancing of the tongue: A model-based learning approach to phonetic targets in coarticulationa).

This article focuses on developing and validating a computational model of anticipatory coarticulation in speech production, based on physiological articulatory data from electromagnetic midsagittal articulography (EMMA). The model conceptualizes the tongue dorsum as a carrier wave and the tongue tip as a modulation signal, reflecting a principal-subordinate structure in articulatory movements. Using a two-layer optimization framework that integrates a physiological 3D articulatory model with statistical learning, the study estimates phonetic targets and model parameters, achieving close alignment with observed articulatory data (average error ~0.18 cm). Subjective listening tests further demonstrate that speech synthesized from the optimized model yields improved naturalness over original or partially optimized articulatory targets. The study acknowledges limitations due to its focus on isolated syllables and suggests future work to extend modeling to continuous, natural speech contexts.