This paper presents an approach for incorporating prosodic knowledge into the language modelling component of a speech recogniser. We formulate features for a maximum entropy language model which capture various aspects of the relationships between prosody, syntax and the spoken word sequence. Maximum entropy is a powerful modelling technique, and well suited to modelling prosodic information. Tests conducted on the Boston University Radio Speech Corpus using this model showed improvements in perplexity, and n-best rescoring results also demonstrated small but statistically significant gains.