In this paper we describe a speech expressivity transformation system giving the opportunity to a user to modify the expressivity of a spoken utterance. Statistical model are learned on a multispeaker expressive database using a Bayesian Network. The acoustic modification of the speech signal is achieved by a phase vocoder technology. The parameters of those transformations are context dependents. They change along the sentence in respect of pragmatic information such as stressing and depending on the phonetic transcription of the text. The system is now working for several acted emotions in french and is used for an artistic purpose dealing with multimedia and cinema.