This paper describes the use of prosodic rules for segmenting speech signal in sentences, words (lexical and grammatical) and putting milestones on their boundaries. The strategy in an automatic speech recognition system uses these milestones for filtering lexical and syntactical hypotheses in order to limit the depth of the lexical and syntactical search. Prosodic parameters (pitch, energy, duration) are estimated over vocalic nuclei and normalized using two types of coding: a) uniform coding, b) variable coding.