Listeners were randomly presented narrow-band filtered noise bursts that varied in filter center frequency from two overlapping, Gaussian-like distributions. Participants mapped these distributions of sounds onto creatures in a video game where they received visual and auditory feedback about their accuracy. Categorization boundaries for each participant were estimated using logistic regression and compared with the optimal boundary from an ideal observer model. The participants appeared to be able to establish near optimal boundaries rapidly and had a remarkable ability to shift these boundaries when the underlying distributions changed - even when these changes were not explicitly signaled. These results suggest that listeners maintain a rather detailed representation of distributional information that is continuously updated during the task. This interpretation is in line with the assumptions underlying many current models of perceptual (statistical) learning in speech perception. However, it is possible to get optimal-like behavior by maintaining a general distributional representation or by using simpler “local” strategies based on only a few of the most recently experienced exemplars. The results will be presented with multiple categorization models, which testify to the difficulty of interpreting claims of distributional learning in categorization. [Work supported by NIH-NIDCD.]