Acoustic echo cancellation (AEC) algorithms are widely used in communication devices. One of the main design problems is double-talk issue. The recursive prediction error (RPE) algorithm based on autoregressive (AR) model has been proven to be more attractive to traditional approaches. However, it adapts too fast in double-talk which may cause the algorithm to diverge. A variable step size (VSS) approach is developed to tackle the problem. On the assumption that prediction error is more than 10

more »

... times smaller than microphone signal when the adaptive filter has converged, the step size is adjusted adaptively according to the energies of near-end, far-end and prediction error signals. The experimental results show that our algorithm, compared with original RPE algorithm, the performance in double-talk is improved more than 8 dB while no any performance losses in single-talk and echo path change cases.