Research On Enhancement Algorithm Of Noisy Mask Speech

The mask worn in underwater operations and various rescue tasks makes the wearer's voice change,and the influence of ambient noise leads to poor voice quality and hinders normal communication.Therefore,the research of face mask speech enhancement algorithm is of great significance to improve the communication quality in all kinds of environments.Speech enhancement algorithm is a way to extract the most clean original speech from noisy speech by suppressing noise interference,so as to reduce noise pollution.For mask speech enhancement,this thesis focuses on the following:(1)Research on the phonetic characteristics of the mask speech.In order to fully grasp the features of mask speech,the short-term energy,short-time average amplitude difference function,pitch period and spectral map are used to compare and analyze the mask speech and air speech,and the time-frequency domain characteristics of the mask speech are summarized.(2)Research on the speech enhancement algorithm of empirical mode decomposition(EMD)and compressed sensing.The EMD wavelet threshold denoising method is used to enhance the speech of air and mask respectively.The results show that the algorithm has an enhanced effect on both types of speech.Under the premise of using wavelet sparse and Hadamard matrix observations,Basis Pursuit algorithm,Match Pursuit algorithm,Orthogonal Matching Pursuit algorithm and Stagewise Orthogonal Matching Pursuit algorithm are used to enhance speech,the experimental results show that Stagewise Orthogonal Matching Pursuit algorithm has the best enhancement effect,however the enhancement is limited.(3)Mask speech enhancement algorithm combining EMD and compressed sensing.The performance of each modal component after noisy speech EMD processing is analyzed,and the effect of speech enhancement which directly removes the intrinsic Intrinsic Mode Function(IMF)with much noise is verified.The sparsity of discrete cosine bases and wavelet transform bases is studied,and the reconstruction performance of random Gauss matrix,part Hadamard matrix and row ladder matrix are also studied.It proposes a speech enhancement algorithm which performs wavelet threshold denoising on high-frequency IMF with noisy signal first,and then performs wavelet transform and Hadamard matrix observation on all processed IMF components.Finally,an improved sparse adaptive regularization compressionsampling matching pursuit algorithm is used to reconstruct the signal.Experimental results show that this algorithm is better than compressed sensing and EMD wavelet thresholding methods.(4)Mask speech enhancement algorithm combined with improved k-singular value decomposition and variational Bayes robust principal component analysis.This thesis analyzes the speech signal recovery performance of three methods: singular value iterative threshold method,accelerated proximal gradient method and Lagrangian multiplier method in low-rank matrix recovery.The variational Bayesian robust principal component analysis is introduced,and the denoising effect is analyzed through experimental simulation.A modified Bayesian robust principal component analysis method based on improved k-singular value decomposition is used to enhance the speech.The results show that the proposed algorithm can effectively enhance the mask speech with noise.