A noise reduction method using linear predictor with variable step-size

A white noise reduction method based on linear prediction has been proposed previously. The linear predictor can reduce a white noise from a noisy speech, because the linear predictor converges such that the prediction error signal becomes white. In the conventional method, the linear predictor is updated by using an algorithm with a fixed step-size for estimating the speech signal. However the optimal step-size, which can reduce the estimation error of the linear prediction, varies with the coefficients of the linear predictor. In this paper, we derive the optimal step-size from the convergence condition of the linear predictor, and propose a variable step-size so that the estimation error of the linear prediction is reduced. Experimental results show that the proposed noise reduction system can improve the noise reduction property in comparison to the conventional one.