Commercial text-to-speech systems are currently based on the so-called corpus-based synthesis technique, which basically consists of retrieving, at the time of synthesis, the speech fragments necessary for the construction of the synthetic utterance of a large voice database previously recorded by a speaker. The design and recording of these voice databases is therefore a key element in the process of building a synthetic voice: its phonetic and prosodic coverage must be rich enough to ensure good results in the process of selection of fragments for synthesis.