Linear Combining of Audio Features for Signal Classification in Ad-hoc Microphone Arrays

Audio signals are often corrupted by signal contributions from competing sources and reverberation in the acoustic environment. In an audio signal classification task these effects introduce a mismatch between test and training data, which decreases the classification accuracy. When multiple sources are simultaneously active and captured by multiple ad-hoc distributed microphones in a room, it is of interest to determine the type of each source based on the captured signal mixtures. Obviously, the microphones closest to a particular source are most suitable for its classification. However, it is not clear how to combine signal features extracted from the microphone signals in an ad-hoc array in order to classify the source signals reliably. In this contribution different data combination strategies are introduced. The resulting classification performance is analyzed based on simulations and audio recordings. When information from microphones within the critical distance of a source is combined with information from the other microphones in the room, a high classification accuracy can be obtained.