The CNET's commercially available text-to-speech system performs an automatic prosodic parsing, based on the detection of a small number of grammatical words. However, the lack of any other syntactic information is a serious source of errors at the supra-segmental level. In this paper, we present a new prosodic parsing algorithm which overcomes that drawback. First, a recursive left-to-right morphosyntactic analysis assigns a grammatical value to each word in the text and transcribes it phonetically. Second, one hundred and forty hierarchized parsing rules divide the message into a sequence of prosodic groups. Finally, prosodic patterns are automatically assigned to each word by queries to a data base of prosodic events. Preliminary tests indicate that such an approach, combining a linguistic processor and a data base of real prosodic features, yields synthetic speech with a high degree of naturalness.