Modeling the Acquisition of Covert Contrast

This paper explores the learnability of covert contrasts (impressionistically homophonous categories that can be reliably distinguished at the phonetic level) through a series of model-based clustering simulations using human production data. Allowing the models to learn both the number and parameters of those categories provides a way to explore the potential stability of category structures. The results indicate that while a statistical learner can be quite effective at inducing covert contrasts, success depends crucially on the number and distributional characteristics of the relevant cue dimensions.