This paper evaluates the performance of two automatic labelling systems for intonation by using their output to predict the position and type of prominences and boundaries labelled in a ToBI transcription of twelve dialogues. It shows that they both model the prosodic information well, and that improvements in modelling are gained when including segmental information about the duration and energy profiles of the utterance. However, after parameter reduction, the features that survive are segmental rather than F0-related.