Research On Speech Enhancement Method Based On Compressed Sensing Theory

Speech is one of the most important information communication tools in human soci-ety,it is not only the most convenient way of communication between people in daily life,but also the most efficient way of interaction between human and computer.In reality,speech signals will be degraded by noise in the process of acquisition and transmission.The presence of noise not only affects the transmission of speech information,but also makes the listener psychologically tired and bored.The purpose of speech enhancement technology is to suppress or eliminate the noise in the noisy speech signal and extract the clean speech signal from the noisy speech signal.After years of development,many speech enhancement algorithms have been proposed,the most classic of which are spec-tral subtraction,subspace methods,Wiener filtering methods,and so on.However,these methods are proposed based on smooth noise and cannot suppress non-smooth noise well.In this paper,we firstly propose a spectral subtraction method based on priori signal-to-noise ratio(priori SNR)estimation,which solves the problem of musical noise in the traditional spectral subtraction method.Then,based on the theory of compressed sensing,we propose a orthogonal matching pursuit(OMP)method based on energy threshold and a dictionary learning algorithm based on joint training samples for speech enhancement,which achieves a better speech enhancement performance.The main work and innovation of this paper are as follows.Firstly,the basic theoretical knowledge of speech enhancement is introduced.For the problem of musical noise in spectral subtraction method,a spectral subtraction method based on prior SNR estimation is proposed in this paper.The analysis shows that the exis-tence of musical noise is mainly due to the inaccurate estimation of noise.The algorithm proposed in this paper transforms the estimation of noise into the estimation of the prior SNR.By introducing an adaptive smoothing factor in the direct decision method to im-prove the estimation accuracy of the priori SNR ratio for each frame,the musical noise in the enhanced speech is reduced and the degree of speech distortion is alleviated.The simulation results show that the proposed method has significantly improved performance compared with the reference algorithm under different types of noise.Secondly,the reconstruction algorithm in compressed perception theory is applied to study the speech enhancement problem.The existing reconstruction methods based on compressed sensing theory aim to reconstruct the original signal with noise,which cannot effectively suppress the noise in speech.To solve this problem,an orthogonal OMP algorithm based on energy threshold is proposed in this paper.Firstly,the whole speech is divided by endpoint detection method,and then the speech frame is reconstructed by OMP algorithm.In the iterative process,an energy threshold is designed to control the iteration times of the algorithm by calculating the energy of the speech components in the noisy speech,so as to realize the adaptive denoising ability and improve the performance of the algorithm.Simulation results show that the proposed algorithm has better performance in noise reduction than other reconstruction algorithms under different noise conditions.Finally,based on dictionary learning theory,a speech enhancement algorithm based on joint training sample dictionary learning is proposed to solve the speech enhancement problem.Existing speech enhancement algorithms based on dictionary learning use clean speech and noise as training samples respectively to obtain learning dictionaries.The dic-tionary constructed by this method can not distinguish signal and noise accurately under some noise condition,which leads to poor performance of noise reduction.To solve this problem,this paper uses noisy speech combined with clean speech and noisy speech com-bined with noise as training samples respectively in the process of learning dictionary to obtain two joint learning dictionary.Two dictionaries are used to obtain the sparse rep-resentation of two signals,and the complementary relationship between clean speech and noise in sparse representation is used to carry out the weighted processing,and finally the enhanced speech is obtained.Simulation results show that the proposed algorithm has better performance than other reference algorithms under different noise conditions.