Complex system identification methods for fast echo canceler initialization

The authors propose a new recursive version of a technique proposed by G. Long and F. Ling (1990) for the initialization of a data-driven echo canceler (DDEC). They prove that the Long and Ling algorithm yields a least-squares solution, and then a new technique is presented which is comparable to the recursive-least-squares (RLS) algorithm. However, the use of a unique training sequence reduces the complexity of the RLS algorithm to that of the least-mean-square (LMS) algorithm. An analysis of the covariance of the estimated weight vector is presented, and simulation results show a remarkable improvement in both convergence speed and steady-state error