Sounds provide substantial information on human activities in an indoor environment, such as an apartment or a house, but it is a difficult task to classify them, mainly due to the variability and the diversity of realization of sounds in those environments. In this paper, sounds are considered as a class of information, to be mixed with other modalities (video in particular) in the design of ambient monitoring systems. As a consequence, we propose a classification scheme aimed at (i) exploiting the specificities of this modality with respect to others and (ii) leaving doubtful events for further analysis, so that the risk of errors is overall minimized. A dedicated taxonomy together with belief functions are proposed in this respect. Belief functions are an adapted way to face the variability of sounds, as they are able to quantify their impossibility to classify the signals when it differs too much from what is known by creating class of doubt. The algorithm is tested on a dataset composed of real-life signals.