Research On Speech Enhancement Algorithm Based On Compressed Sensing

Speech enhancement has been used to restrain noise, improve the perceptual quality. It is one of the most important ways to promote voice apply system's performance in noise environment. With the rapid development of mobile communication network, more and more kinds of mobile communication equipments have been used, when one side of the mobile communications in a pandemonium condition, the quality of voice communication will be affected by different degree, the study of the speech enhancement become necessary. This paper researched the speech enhancement algorithm based on compressed sensing method. The main work is as following:First, making use of the speech signal's de-correlating property in the frequency domain, exchange noise with speech by DFT transform to get the corresponding power spectrum. Use dictionary learning algorithm to get the corresponding dictionary of speech and noise, and then their respective sparse coefficient of power spectrum can be obtained by sparse transform. Reset the sparse coefficient according to the noise dictionary to zero and the enhanced speech power spectrum can be got through compression perception recovery algorithm. Finally, we can get the time-domain voice enhanced by DFT transform.Second, this research trains the dictionary by recursive least squares dictionary learning algorithm. Due to the power spectrum is negative, make non- negative constraints on it. The simulation results show that this algorithm can improve dictionary training speed without reducing recovery accuracy compared with the DCT orthogonal and K-SVD dictionary learning method.Last, this paper puts forward with the OMP algorithm(OMP algorithm based on termination criterion of noise power spectrum estimate, NPEC-OMP) which is based on criterion of noise power spectrum estimate termination to get the sparse coefficient of speech and noise power spectrum, namely using the estimated noise power spectrum to constraint residual error when the signal sparse representation is obtained by the dictionary. Compares the method in this paper with spectral subtraction algorithm and the subspace method with different noise sources and SNR(Signal to noise ratio) through simulation. The simulation results reveal that compared with the traditional de-noising methods, the speech enhancement effects of the proposed method in this paper under different noise conditions are all improved and the proposed method can adapt to greater SN R range.