A pair learning algorithm for multi-channel blind source separation and its convergence analysis

摘要

A pair learning algorithm is proposed -in this paper. A feedback weight c{sub}(ij)(n) from the jth output y{sub}j(n) to the ith observation x{sub}i(n) is updated using only the ith and jth outputs. Then, this method is called "Pair Learning Algorithm". This algorithm is compared with a global learning algorithm, which use all outputs in updating the feedback weights. The proposed method is a simplified version of the global one, then, the number of computations is reduced and a learning process is simplified. Separation performance of the global method is slightly better than the pair learning. However, their difference is small. Furthermore, the following properties are analyzed: Order of the separated signal sources is dependent on power of the signal sources included in the observations. Magnitude of the separated signal sources are uniquely determined in the pair learning algorithm after convergence. Squashing nonlinear functions have effects of stabilizing the learning process. Simulation using voices, white noise and music signal sources are demonstrated.