Research On Speech Denoising Technology Based On Compressed Sensing

Twice rate of the signal bandwidth is required to sample the signal by using the traditional digital signal processing method based on Nyquist sampling theorem. Thousands of samples will be obtained by digital sampling massive data to be processed, which brings great pressure to the already scarce frequency resources. Meanwhile, samples offen need to be compressed before transmission, results in wasting lots of sampling resources. As a novel compression sampling technology proposed in recent years, compressed sensing is no longer limited by the bandwidth of signal, it can sample signal with much lower sampling rate than Nyquist’s while achieving compression of signal, which significantly saves samples and frequency resources. Compressed sensing technology is a sparse signal processing technology. Since speech signal is approximately sparse, we can combine compressed sensing with speech signal. Based on the sparsity of speech, This thesis studies speech compressed sensing system, especially focuses on studying speech denoising technology based on compressed sensing. It is a key issue must be solved for the practical application of compressed sensing technology.Firstly, the basic theory of compressed sensing is described in detail. After verifying the sparsity of speech signal, the basic theory of speech compressed sensing is be analyze: sparse transform, projection matrix and reconstruction algorithm. While in reality, speech is noisy, so we focus on studying effects of noise on each part of speech compressive sensing system and simulating it.Secondly, given the different sparsity in DCT domain between speech and white Gaussian noise, the row echelon matrix which performs well in speech compressed sensing is used as the measurement matrix. Considering the shortcomings of traditional OMP algorithm: it is difficult to achieve speech enhancement and the iterative speed is slow, we propose a modified OMP algorithm to controls the reconstruction stage and iterations of OMP by setting correlation threshold and energy threshold, effectively suppresses the noise, improves output signal to noise ratio and shortens reconstruction time.Finally, an atom selection method is used to reduce the residual noise produced by sparse representation denoising; Given the bad performance of BPDN algorithm which set parameter ? to experience value, a modified BPDN algorithm that can adapt ? to the optimum with input signal to noise ratio is proposed; Then we combine atom selection method with MBPDN algorithm, design a speech denoising method based on over-complete dictionary. The experimental results show that the proposed system has a good suppression effect on both white noise and colored noise.