Non-stationary source separation and system identification for MIMO-AR mixtures using HM-GMM

Abstract

A technique for non-stationary source separation is proposed for multiple-input multiple-output auto-regressive (MIMO-AR) mixtures. Hidden Markov Gaussian mixture model (HM-GMM) is employed in this paper to represent source model. Expectation-maximum algorithm is used to achieve MIMO-AR model identification, and then blind source separation (BSS) is achieved by matrix joint diagonalization. Simulation results show its effectiveness, not only for system identification, but also for BSS.