Determining the Number of Sources with Diagonal Unloading in Single-Channel Mixtures

Abstract

A required step to realize blind source separation of a signal is to determine the number of sources building up the mixture. Generally, the assumption of equal number of sources and mixtures is imposed which becomes invalid in single-channel mixtures. Therefore, in order to apply conventional solutions, a single-channel mixture signal can be converted into a pseudo-multi-channel signal. The empirical mode decomposition (EMD) method can be used for this purpose while revealing the hidden modes of the signal. In this paper, the number of sources building the single-channel mixture is determined based on the eigenvalue information of the intrinsic mode functions extracted by the EMD method. In this context, the method based on the ratio of adjacent eigenvalues (RAE) is shown to provide satisfactory results. Moreover, the extensions of the RAE method such as diagonal loading (DL) have been already offered for signal detection problems. However, it is presented here that the DL method may not be suitable for mixtures having more than three source signals. Accordingly, a modification combining both methods is proposed. Evaluations for various synthetic signal mixtures, real-world audio and speech samples are carried out to demonstrate the better performance of the provided method.