With an increasing awareness of the needs of the disabled comes a major surge in research into ways of helping them through technology. One specific area is the use of computer-generated speech output to overcome the disability of being non-verbal. Many such `communicators' are presently available, but they are only able to converse at a relatively low rate. This paper describes a method of improving rates of prosthetic conversations through the use of word prediction in conjunction with grammatical recognition. A grammar is constructed consisting of a set of word sequences each element of which consists of a literal word or word type. Following the entry of a word, the prediction system consults a word type dictionary in association with the grammar, to provide a short list of the most likely successors. This list is considerably more accurate than that obtained from systems with no grammatical knowledge. The paper also suggests how speech recognition can be used to improve the prediction facility by analysing the other side of the conversation to provide semantic information.