Naturalness in synthetic speech is to a large extent determined by how well the system models the variability found in human speech. Good models of variability are now emerging, and this paper describes how variability of several different types is incorporated into SPRUCE - a high-level text-to-speech synthesis system. The synthesiser is carefully engineered according to the requirements of a recent computational model of speech production. The resulting voice output illustrates the usefulness of well motivated theory in speech synthesiser design.