Internal representations of the environment may alleviate sensory processing by modulating the gain of incoming information through predictive coding mechanisms. Previous research has has shown that contextual predictions modulate neural signals and that pupil diameter increases during unpredictable audio events. Here, we investigated whether contextual speech predictions modulate pupillary dynamics. A neural-network language model was trained on the speech data form the TED-LIUM corpus to generate phoneme-level predictions manifested as uncertainty and surprise. Video of subject's pupils was obtained during speech-listening. Then, pupillary dynamics were modeled as a linear combination of speech-audio content and contextual speech predictions. However, surprise and uncertainty regressors did not improve modeling of the pupil signal. Further work will assess simpler predictive coding models against the speech-audio model, with the eventual goal of relating pupil dynamics with predictive neural signals.