A priori SNR estimator based on a convex combination of two DD approaches for speech enhancement

摘要

One of the most used techniques for estimating the a priori SNR in many speech enhancement systems is the decision-directed (DD) approach, which is famous for the good performance in reducing musical noise with low complexity calculation. However, the conventional DD approach with fixed smoothing factor suffers from the inherent tradeoff between speech distortion elimination and musical noise reduction. In order to overcome such deficiency, a new approach is proposed by using a convex combination of two DD methods with different smoothing factors, for the purpose to obtain a modified a priori SNR estimate with small speech distortion and musical noise residual. The proportion of either DD method to the convex combination is balanced by a mixing parameter, whose value is restricted to the interval [0, 1] and can be optimally computed based on the minimum mean square error operator (MMSE) criterion. The property of our proposed method is tested with a Wiener filter-based noise reduction system, and the result shows the superiority to some other a priori SNR estimators in different noise conditions and SNR levels.