On integrating speech coding functions into echo cancelling filters with decorrelating properties

The normalized LMS (NLMS) algorithm has been successfully used in many system identification problems. However, the NLMS algorithm is known to exhibit low convergence speed especially when the input data covariance matrix is ill-conditioned. In this paper, we consider a sub-optimal implementation of the affine projection (AP) algorithm based on a prewithening mechanism, which renders the convergence characteristics less sensitive to the coloring of the input signal spectrum than is the case for the NLMS algorithm. Comparisons with the AP algorithm are given to validate our approach. Implementation details are discussed in the context of hands-free telephony where echo cancelling and speech coding algorithms are integrated on the same DSP board.