Improved a Priori SNR Estimation for Speech Enhancement Incorporating Speech Distortion Component

The well known decision-directed (DD) approach drastically limits the level of musical noise, but the estimated a priori SNR matches the previous frame rather than the current one. Plapous introduced a novel method called two-step noise reduction (TSNR) technique to refine the a priori SNR estimation of the DD approach. However, the performance of this method depends on the accurateness of the estimated speech in its second step. In this paper, we propose an improved approach for the a priori SNR estimation in DCT domain with two steps like the TSNR method. While in the second step, considering the two state components of the estimation error between speech signal and its estimation, the speech distortion component and residual noise component, we make the estimated speech subtracted by its speech distortion as a refined estimation for the clean speech signal. Because the speech distortion component is offset, the estimated a priori SNR is more accurate. A number of objective tests results show the improved performance of the proposed approach.