Blind separation of sources based on their time-frequency signatures

Blind source separation based on spatial time-frequency distributions (STFDs) has been previously introduced. This method provides improved performance over blind source separation methods based on second-order statistics, when dealing with nonstationary signals that are localizable in the time-frequency domain. In the STFD method, the covariance matrix is first used to whiten the signal vector, then the unitary matrix is estimated using high signal-to-noise ratio (SNR) time-frequency points. This paper modifies the STFD method by performing both whitening and estimation steps using STFD matrices. The eigenvectors of the signal subspace obtained from a properly selected STFD matrix are more robust to noise than those obtained from the covariance matrix and, therefore, are more appropriate to use, particularly for low SNR environments. Further, for small array apertures and high number of arrivals, the STFD whitening matrix can be used as means to reduce the number of signals considered by the blind source separation algorithm