Split filter structures for LMS adaptive filtering

Abstract

In this paper, the performance of a split adaptive filter in either a parallel or a serial form for non-white inputs is investigated. A parallel split structure is constructed by using two linear phase filters connected in parallel while for serial split it is configured as a cascade of two transversal filters. The adaptation characteristics of the well-known LMS algorithm has been shown to be governed by the eigenvalue spread of the input correlation matrix. By adopting the split structures, we illustrate that the eigenvalue ratios of the associated covariance matrices can be reduced thereby giving rise to a faster convergence speed. The parallel and serial split adaptive filters are examined for both joint process estimation and linear prediction. A new type of linear predictor is formed by combining both split methods and its performance for speech analysis is studied. Simulation results are included to validate the superiority of the proposed split filter structures in improving the rate of convergence for LMS adaptation.