Improved cost functions for blind source separation based on linear prediction

Some improved cost functions based on linear prediction for blind source separation (BSS) are proposed in this letter. For uncorrelated sources, our analysis shows that each proposed cost functions is a generalized Rayleigh quotient, so the optimization can be converted to a generalized eigendecomposition of a corresponding matrix pencil. The proposed cost functions are legible and simulation experiments show that they have better performance than that in the literature.