This chapter focuses on the processing of the major prosodic unit, called intonational period, in Rhapsodie. We explain why the grammatical concept of sentence cannot be taken as a reference for the segmentation of unelicited speech into major prosodic units. Rather we show that in order to segment an utterance into major prosodic units in the context of an emergent data-driven approach, it is necessary to highlight accurate acoustic cues that produce perceptual effects of demarcation between two units. First, we present the Analor model developed to segment discourse into intonational periods, in particular the different parameters selected for segmentation, and how they interact. Second, we illustrate the choices made to process fillers, disfluencies, and also to resolve problems of segmentation derived from overlaps and backchannels in dialogic samples.