In a dialogue, there are at least two sorts of boundaries between discourse units. One type of boundary signals the end of a topical unit; another type of boundary the end of a turn at talk. These two do not necessarily coincide, as a speaker may wish to a new topical unit without wanting to be interrupted by his interlocutor. In order to test whether prosodic cues can differentiate unambigously between topic and turn boundaries, a series of production experiments was set up in which topic-finality and turn-finality were varied independently, and in which visual and non-prosodic verbal cues could not be used. In the most complex condition, the speaker had to give clear cues for topic finality, while not prematurely losing the floor. In this condition, speakers avoided using low tones at turn-internal topical boundaries, reserving them to signal turn-final topic boundaries. When listeners were confronted with portions of the description taken out of their contexts, they could reliably differentiate between turn-final and non-turn-final topical units. Interestingly, when the final parts of a topical unit were removed, listeners could still discriminated between turn-final and non-turn-final expressions, apparently basing themselves on other, more global, prosodic cues. This holds similarly for both minimally and maximally incomplete units.