Blind detection of the number of communication signals under spatially correlated noise by ICA and K-S tests
1. Introduction Estimating the number of sources provides useful information for signal processing applications, such as blind source separation (BSS) in the frequency domain . It is well known that the number of dominant eigenvalues of the spatial correlation matrix corresponds to the number of sources [2,3]. However, it is difficult to distinguish dominant eigenvalues from the other eigenvalues in a reverberant case as shown in Sect. 3. This difficulty has already been pointed out in , where they propose the use of support vector machines (SVM) to classify eigenvalue distributions and determine the number of sources. However, the SVM needs to be trained beforehand and experimental results were provided only for 1-or 2-source cases. This letter first discusses the problem of the conventional eigenvalue-based method when applied to a reverberant condition. Then we propose a new approach for estimating the number of sources that employs independent component analysis (ICA). Experimental results show the characteristics of the proposed approach compared with the conventional eigenvalue-based method.