In this paper we investigate the usefulness of a probabilistic context-free grammar (PCFG) for assigning prosodic structure to unlabelled text. We develop and train a grammar for experiments on German, utilising prosodic non-terminal categories such as phi-phrases. The PCFG is evaluated on test data and by human blind labelling. The statistical prosodic rules can be used in a text-to-speech synthesis system for determining the location of prosodic breaks.