Due to the low computational complexity and acceptable ability in reducing musical noise effect

the decision-directed (DD) approach is widely used for estimating the a priori signal-noise-ratio (SNR) in many speech enhancement systems. However

the DD approach suffers from the problem of time delay and the performance is very sensitive to the fixed smoothing factor. Firstly

the performance of DD approach in musical noise reduction as well as speech distortion attenuation are analyzed using actual speech and noise data

and the boundary values of smoothing factors are presented in view of the analyzed results. Then

a novel algorithm is proposed

in which two DD approaches with different smoothing factors are probabilistically combined in an attempt to put together the best properties of them. The contribution of either DD approach to the combination is automatically adjusted in accordance with the speech absence probability

which can be computed using the complex Gaussian model and soft decision technique. Experiments are carried out in different noise and input SNR conditions

and the results demonstrate that the proposed algorithm can significantly outperform the popular methods for estimating the a priori SNR.