Two text-to-speech systems for Arabic language have been realized in LEESA. Intelligibility of synthetic speech is satisfactory, however prosodic informations (intonation and rhythm) necessary for improving the naturalness remain rudimentary. In this paper, we will report on the adopted approach in modeling the variations of the vowels durations in Arabic speech. For each phonetic segment, a special coding includes lexical, phonetic information and their durations. Interrogation of database is realized by a procedure of the type "searching by key-words" implying a combination of the functions "And, Or, But, Neither". The interrogation is done in a sequential way and allows an automatic and interactive processing in getting statistics. Using the statistically significant factors, a duration model of Arabic vowels is proposed. The duration prediction using this model shows that the root mean square error between prediction duration and duration found in the data-base is equal to 14.74 ms.