Towards a Speech Enhancement De-noising Algorithm Based on Cepstrum Smoothing Prior SNR Estimation

The accuracy of prior Signal to Noise Ratio (SNR) estimation directly affects the performance of single-channel speech enhancement de-noising algorithm. In the single-channel enhancement algorithms based on short-time amplitude spectrum estimation, the gain is usually described as function about priori and posteriori SNR. The decision-directed (DD) method is widely applied in prior SNR estimation. Based on DD method, the smoothing factor was dynamically adjusted in accordance with existing probability of speech frame. The modified Two Step Noise Reduction (TSNR) algorithm was used to arrive at estimation of prior SNR. The cepstrum smoothing estimation method was introduced to reduce background noise on the basis of adaptive smoothing algorithm. The experiment comparison result shows that the proposed cepstrum smoothing estimation method may correct fluctuation variance caused by misjudgment and reduce remaining noise to certain extent due to different smoothing with different coefficients in the cepstrum field.