Using MBROLA database, it is possible to generate pretty realistic human voice. However, tedious work is necessary to tune individual sentences. Thanks to an automatic process using a phonetic dictionary, we are able to generate instructions for the MBROLA engine which then produces decent non-robotic speech.