Research And Improvement For Several Speech Enhancement Algorithms And De-noising

The goal of speech enhancement is to withdraw the pure primitive speech from the noise speech,the actual speech enhancement mainly have two targets: eliminateing the background noise and improving the speech quality. We detail the speech enhancement technique in the circumstance of adding noise. First of all,we provide basic theory of speech signal processing,which is foundation of the research and implement of speech enhancement,and then we introduce traditional speech enhancement method such as spectral subtraction, wavelet transform and subspace method,finally,we compare the advantage and shortcoming of these methods.We focus on research of the following aspects :a improved method is provided responsing to a defect of the spectral subjection theory and the experiment shows the method can hold the features of weak components effectively; we present a improved algorithm based on characteristic of energy distribution and division method, the experiment shows that noise can be restrained and image can be effectively decomposed ;based on wavelet transform and the second order statistics of the signal, a kind of adaptive blind source separation algorithm is presented, the eigenvector matrices of the mixed signals and the whitened signals are obtained by using the point property of a special cost function, simulation shows that the algorithm can get good separation performance; to overcome the difficulty of soft threshold algorithm, a improved threshold is presented, which improves the speech quality more effectively; we also develop adaptive wavelet transform based on the lifting scheme, making Bernstein filter predictor adaptively to match a desired signal by adaptive criteria and then apply it on the signal de-noise, the experiment shows that the SNR of the method is higher than normal WT, the advantages of lifting scheme lie on its flexible design and compute in-place; with masking perceptual model to calculate the noise masking thresholds parameter, we can update the estimated noise instantly and get a compromise proposal between speech distortion and remain noise, experimental result shows that the new approach realizes better enhancement effect than MMSE method.