Detecting and tracking broad sound classes in audio documents is an important step toward their structuration. In the case of complex audio scenes, such as the sound track of a TV broadcast, one problem is that several classes of sound maybe present simultaneously. It is therefore important to detect such superimposed events. Most methods would necessitate to estimate a model for each combination of sound classes that is to be detected, which is intractable in practice since it requires a lot of manual labelling. In this paper, we propose and compare several approaches to detect simultaneous events using only the models of the base classes we are interested in. Two main approaches are compared: model combination and binary hypothesis tests. The results show that the best results are obtained with the model combination approach.