To produce synthetic speech which sounds natural, it is essential to segment longer sentences into breath groups and to assign appropriate intonation contours to these segments [Ainsworth, 1973]. To this end knowledge of the syntactic role played by each word in a sentence will be necessary in a complete text-to-speech system. Though the addition of part-of-speech information to an existing pronouncing dictionary requires little additional space, other approaches such as connectionist text-to-speech systems [Ainsworth & Pell, 1989] require alternative means for syntactic analysis. Here we report on a connectionist classifier, trained to predict parts of speech from orthography. We have found that the scores obtained for unseen words, with networks of less than 5000 weights, are high enough to provide a storage-efficient alternative to a grammatical dictionary. These scores compare favourably with those of a simple set of rules.