Integrating prediction errors at two time scales permits rapid recalibration of speech sound categories

Abstract

Speech perception is assumed to arise from internal models of specific sensory features associated speech sounds. When these features change, the listener should recalibrate its internal model by appropriately weighing new versus old evidence in a volatility dependent manner. Models of speech recalibration have classically ignored volatility. Those that explicitly consider volatility have been designed to describe human behavior in tasks where sensory cues are associated with arbitrary experimenter-defined categories or rewards. In such settings, a model that maintains a single representation of the category but continuously adapts the learning rate works well. Using neurocomputational modelling we show that recalibration of existing "natural" categories is better described when sound categories are represented at different time scales. We illustrate our proposal by modeling the rapid recalibration of speech categories (Luttke et al. 2016).