A method is described for automatically segmenting a database of diphones for speech synthesis purposes. This involves using hidden Markov models to find phoneme boundaries and then a spectral mismatch minimisation algorithm to choose the precise diphone boundaries.