An adaptive control method for regression coefficient in frequency domain non-linear echo suppressor

An echo canceller, which consists of a linear echo canceller and a non-linear echo suppressor, has been proposed for mobile phones. This approach provides an efficient nonlinear echo suppressor with low computational complexity. In this method, the non-linear echo is estimated in the frequency domain from the linear echo replica. A ratio of the non-linear echo and the linear echo replica is called `Regression Coefficient'. In this approach, it is very important how to estimate and how to update the regression coefficient by using the residual signal after the linear echo canceller, which includes the residual non-linear echo and the near-end signal, and the linear echo replica. In this paper, we propose a new adaptive control method for the regression coefficient. The ratio of the residual signal and the replica quickly changes along frame, which is an interval of FFT. Therefore, first the average of the ratio is estimated. Second, the average is amplified in order to suppress the residual non-linear echo. In the proposed method, the regression coefficient is not updated in the double talk intervals, because effects of the near-end signal is large. The double talk intervals are detected based on the correlation coefficient between the linear echo replica and the microphone input signal, which includes the linear and non-linear echo components and the near-end signal. Simulation results show the proposed method can improve echo reduction and the segmental SNR by 13 dB and 1.5 ~ 2.5 dB, respectively.