Lessac Technologies has developed a technology for concatenative speech synthesis based on a novel approach for describing speech in which expressivity, voice quality, and speaking style are fundamental. The main aspect of our system is that instead of traditional phonetic symbols, we use a much more fine-grained and richer set of entities called Lessemes to describe speech and to label units, which allow a richer and more precise characterization of speech sounds. The front-end part of our synthesizer translates plain input text into a sequence of these units by syntactic parsing and applying a set of rules developed from expertise. We use a Bayesian method to obtain a particular trainable mapping from linguistic and prosodic features encoded in the Lessemes to a trajectory in the acoustic parameter space. Unit selection consists of selecting the best candidate units from a data base to match them to the target trajectory, while minimizing discontinuities between them.