When precise expectations lose: Precision and Competition Among Hidden Causes in Voice Recognition

Abstract

Perception is often described as Bayesian inference, where precise expectations are assumed to dominate perception by pulling estimates strongly toward their mean. However, this effect may reverse when the hidden cause of the sensory input is unknown and must itself be inferred among competing alternatives. We tested this prediction in a voice-recognition task in which listeners learned three voices of speakers that differed in precision, defined as the inverse variance of the acoustic features of their voices. When classifying ambiguous utterances, listeners across five experiments systematically favored speakers with lower-precision, i.e., more variable voices, an effect that was strongest under high sensory ambiguity and amplified by explicit knowledge of speaker variance. Computational modeling showed that listeners tracked and used idiosyncratic, speaker-specific precision estimates that increased with learning accuracy and were stable across sessions. These findings show that precision does not simply determine the strength with which perceptual estimates are drawn toward a prediction; when the hidden cause must be inferred among competing alternatives, it also determines which prediction is favored. Our results identify precision as a critical determinant of perceptual inference when multiple hidden causes provide competing predictions about a stimulus.