Incoming raw sensory content is diverse and dynamic, posing a challenging computational problem for brain processing. Internal representations of sensory contexts may alleviate this processing; these models suggest that the gain of incoming information depends on whether stimuli are expected or unexpected. Mechanistically, this is referred to as predictive coding. Predictive processing manifests in various physiological contexts. For example, increases in pupil diameter during surprising events is hypothesized to modulate the gain on unexpected stimuli, and hierarchically organized neural oscillations are thought to mediate top-down relaying of predictive information. Here, the roll of physiological dynamics in the pupil and brain in response to contextual speech predictions are investigated. A neural network model was trained on human speech to generate speech prediction parameters as a proxy for auditory processing. The dynamics of physiological responses were modeled using speech features and prediction parameters in order to determine the timescale of their dynamics. Pupil responses showed increased diameter during highly surprising events, and neural activity was better predicted at varying timescales depending on the cortical region's position in the processing hierarchy.