Blind separation of nonstationary sources in noisy mixtures

Abstract

A new method of blind source separation is presented that is not sensitive to additive white noise. The method exploits the nonstationarity and temporal structure of sources. The method only requires multiple time-delayed correlation matrices of the observed data at several different time-windowed data frames to estimate the mixing matrix. The implementation using joint diagonalisation (JD) is described. Simulations verify the high performance of the proposed method, especially in a low SNR environment.