A Robust NLMS Algorithm for Acoustic Echo Cancellation

摘要

When an adaptive filter is applied for an acoustic echo canceler, realization of robustness in the presence of double talk is an important subject. In this paper, based on M-esti-mation, a robust statistical technique, a cost function expressing the normalized least-mean-square algorithm (NLMS), is modified to derive the gradient-limited NLMS (GL-NLMS) algorithm. This technique makes use of the statistical characteristic that the ratio of the error and reference signal levels has different distributions during single talk and double talk. The gradient used for updating in the NLMS algorithm is limited to a nonlinear form, so high robustness against double talk is achieved.