Perception balances what we sense with what we expect. To maintain accuracy, our brains must continuously update sensory expectations when predictions don't match reality. To investigate this, we trained mice on an auditory detection task with hidden switches in stimulus probability across trials. Mice dynamically adjusted their decision thresholds, tracking these changing sound probabilities. We monitored dopamine in the tail of the striatum (TS), a region at the interface of sensory processing and action selection that receives direct auditory cortex inputs. Given mesolimbic dopamine's established role in reward prediction errors, we hypothesized it might perform analogous computations for sensory predictions. We found large dopamine transients to sensory inputs that scaled with signal intensity and varied inversely with learned sensory expectations---hallmarks of prediction error encoding. To formalize this, we developed an online Bayesian inference model that updates sensory beliefs about stimulus probability based on recent experiences. The updating rule adjusts expectations based solely on sensory observations, independent of outcome, to better predict future inputs. The model's prediction errors---differences between expected and actual inputs---explained both the animals' behavioral adjustments and TS dopamine responses. These findings provide a unified framework for TS dopamine as a sensory surprise signal, encoding prediction errors that guide responses to unexpected events---from salient threats to routine sensory changes.