People who have lost the use of their voice through disease or accident often require a prosthesis to help them carry out conversations with those around them. Unfortunately, the drawback of such systems is that they are prohibitively slow and the user can often find themselves getting left behind in a conversation. Prediction systems which attempt to guess what the user will say next can help improve this speed, but these also suffer disadvantages. Inaccurate modelling can result in the prediction of unusable words unnecessarily increasing cognitive load on the user. A system is required which will only produce syntactically correct predictions. Windmill is a MS-Windows system which aims to fulfil this requirement using an augmented phrase structure grammar to parse semi-complete sentences. The system is designed as an expert system with a rule base providing possible sentence-completion sequences from which parts of speech for the next word can be deduced. A type-tagged lexicon is then consulted to obtain words which fit these types. A statistical method, based on frequency and recency of usage, is applied to this output to produce an ordered prediction list the top five words of which are offered to the user. Initial results from Windmill indicate a potential of up to 63\% keystroke saving using a simulation test platform. This shows that syntactic processing is a viable option in efforts to improve the communication rate of non-vocal people.