A multi-sensor approach for real-time detection and classification of impact sounds

We present a method for real-time detection and classification of impact sounds - relying solely on spatial features - that exploits the difference in the location of each impacted structure. Using a compact sensor array, we formulate the classification problem in terms of an undetermined source separation process where we assume that the linear mixing model can be learned through a training phase. The recovered source amplitudes are exploited for estimating the source activity in time, and the detection and classification decisions are derived based on simple energy criteria. Experimental results with two sensors demonstrate the efficiency of the method in an application scenario which considers the use of a simple object as a real-time control interface for triggering a percussion synthesizer.