Adaptive speech enhancement based on wavelet in high noise environment

An Adaptive speech enhancement method in high noise environment is proposed in this paper. The whole algorithm is based on wavelet transform and the improved Stein's Unbiased Estimate of Risk (SURE). This work improves the basic wavelet shrinkage method from shrinkage function and enhancement algorithm. Firstly, the detection algorithm is introduced to estimate the noise feature and improve the speech intelligibility. Secondly, we present a simple shrinkage function that is continuous like soft and more stable than hard. Finally, we employ Least Mean Squares (LMS) to estimate shrinkage threshold, which can adaptively track the changes of the system in real time and continually seek the optimum in some statistical sense. These modifications enable the system to handle colored and nonstationary noises. To evaluate the system performance, we have employed plentiful speech data that are obtained under practical environment. Subjective and objective evaluations show that the proposed system improves the performance and speech intelligibility in noisy environment.