Speech enhancement using backward adaptive filtering algorithm: Variable step-sizes approaches

This paper addresses the problem of speech quality enhancement and acoustic noise reduction by adaptive blind source separation algorithms. Recently, several two-channel forward-and-backward structures have been proposed for acoustic noise reduction based on normalized least-mean-square (NLMS) algorithm. The backward structure presents the good performance compared with its direct version (forward). But the main drawback of the backward algorithm is its poor performance in steady state regime when the fixed step-size is selected large. In this paper, we propose two new versions of this algorithm by introducing simple variable step-sizes approaches, which allow improving both the steady state values. The proposed modifications are based on efficient estimation of the optimal step-sizes of the cross-coupling filters. To confirm the good performance of the proposed variable step-size algorithms, we have carried out several simulations in very noisy situation.