Blind source separation using joint canonical decomposition of two higher order tensors

摘要

A two-stage-type algorithm is presented for blind source separation in the overdetermined instantaneous mixture case. The algorithm accomplishes two tasks: blind identification for estimating the mixing matrix and source estimation for recovering the original source signals with the identified mixing matrix. In this paper, we focus on the former task. A new mixing matrix identification method, which is based on the joint canonical decomposition of two third-order tensors, is proposed. Generalized singular value decomposition technique is used to perform the joint canonical decomposition. The merits of the proposed algorithm include the robustness to noise and superior performance compared with the classical blind source separation algorithms. Simulations for speech source separation are conducted to demonstrate the effectiveness of the proposed algorithm.