A second-order feature window method for blind separation of speech signals corrupted by color noise

A second-order feature window (SOFW) method is proposed for blind separation of speech signals corrupted by color background noise based on the short-time stationarity property of speech signals. At first the new prewhitening algorithm is developed to remove the effect of color noise, then the prewhitened data is partitioned continually by feature window with the length equal to the cycle of the fundamental tone of speech signals, after that the speech signal can be separated blindly by estimating the Givens rotation parameters based on the joint approximately diagonalization theory. This novel second order statistics-based method exploits the temporal structure inherent in speech signals, it is simple and can be used as the preprocessing stage for speech recognition systems. Simulation results show that it outperforms the existing typical blind source separation algorithm for speech signals contaminated by color noise.