High-level sensory neurons encoding natural stimuli are not well described by linear models operating on the time-varying stimulus intensity. Here we show that firing rates of neurons in a secondary sensory forebrain area can be better modelled by linear functions of how surprising the stimulus is. We modelled auditory neurons in the caudal lateral mesopallium (CLM), of adult male zebra finches under urethane anaesthesia with linear filters convolved not with stimulus intensity, but with stimulus surprise. Surprise was quantified as the logarithm of the probability of the stimulus given the local recent stimulus history and expectations based on conspecific song. Using our surprise method, the predictions of neural responses to conspecific song improved by 67% relative to those obtained using stimulus intensity. Similar prediction improvements cannot be replicated by assuming CLM performs derivative detection. The explanatory power of surprise increased from the midbrain through the primary forebrain and to CLM. When the stimulus presented was a random synthetic ripple noise, CLM neurons (but not neurons in lower auditory areas) were best described as if they were expecting conspecific song, finding the inconsistencies between birdsong and noise surprising. In summary, spikes in CLM neurons indicate stimulus surprise more than they indicate stimulus intensity features. The concept of stimulus surprise may be useful for modelling neural responses in other higher order sensory areas whose functions have been poorly understood.