We propose in this paper an entirely probabilistic approach to sleep analysis. The analyser uses features extracted from 6 EEG channels as inputs and predicts the probabilities that the sleeping subject is either awake, in deep sleep or in rapid eye movement (REM) sleep. These probability estimates are provided for different temporal resolutions down to 1 second. The architecture uses a “divide and conquer” strategy, where the decisions of simple experts are fused by what is usually refered to as “naÿve Bayes” classification. In order to show that the proposed method provides viable means for sleep analysis, we present some results obtained from recordings of good and bad sleep and the corresponding manual scorings.