Classical speech synthesis systems either concatenate diphone-like tabulated patterns or reconstruct speech parameters according to pre-defined rales. Both techniques show drawbacks: the former lacks flexibility while the latter is highly time-consuming to built.
We propose an intermediate technique using structured segments: segmental units are still resorted to, but they are automatically analysed in terms of a set of spectral targets, a temporal decomposition pattern and a parametric glottal excitation. Structured segments can then be handled by rules. They also supply a valuable material which can be referred to, for building gradually a reconstructive synthesis system.