In this paper, we investigate the problem of automatic audio surveillance. This aspect of the surveillance, which extends the more investigated area of video surveillance, can be very informative to solve many problems in real situations. Similarly to video surveillance, also in this case it is necessary to build a background (BG) model, so that it is immediate to discover foreground (FG) events. To this end, we first introduce the concepts of audio BG and FG in an automated surveillance scenario. Subsequently, we propose a novel audio BG system able to build in real time an adaptive model of the audio scene BG, and to promptly detect unexpected FG auditory events. The method is based on the probabilistic modelling of the audio data stream using separate sets of adaptive Gaussian mixture models, working on the audio frequency spectrum. This approach is also characterized by the use of only one microphone and on-line functioning, so that it can be directly used in real situations, also to support a video surveillance system. Preliminary results show the effectiveness of the approach to discover different FG audio situations.