Source Number Estimation and Clustering for Underdetermined Blind Source Separation

Much research has been undertaken in the field of blind source separation (BSS) and a large number of algorithms have been developed. However, most of them assume that the number of sources is known. In this paper we present an algorithm to estimate the number of sources in the (over-)determined and underdetermined case. We call this algorithm NOSET (Number of Sources Estimation Technique). We start from a description of the BSS problem, give a short overview of the so-called observation vector clustering algorithm and then present our approach. It is based on direction-ofarrival (DOA) estimation from reliable time-frequency points and a clustering of the DOA estimates. The estimated DOAs can be used to recover the source signals by performing a nearest-neighbor classification of the observation vectors instead of the conventional kmeans clustering procedure which is sensitive to the choice of initial centroids.