Audio Event Detection Using the Mapping Segment on the Dictionary in Sparse Representation

Audio event detection (AED) is addressed by using a segment mapping on the NMF dictionary in the sparse representation. One problem with dictionary methods is the lack of controls in the decomposition process of the input signal, so the process yields some unstructured sound pieces that are not the valid audio events. We proposed an algorithm which uses sparsity constraint and beta-divergence to decompose the input segments into the predefined dictionary atoms instead. Here, the sparsity control in each segment decomposes it into a linear combination of basis vectors thereby the segment is approximated into a hypothetical audio event. This method is applied to the recognition of variety live official sound events and has promising results.