Research On Improved NMF Speech Enhancement Method

Voice is the carrier of information exchange.In real life,pure voice is often interfered by noise,such as background noise,other unrelated voice signals and reverberation in closed environment.Noisy speech will make people’s subjective auditory quality decline,and then the intelligibility of speech content is damaged.As a branch of digital signal processing,speech enhancement is an important means to deal with this kind of damaged speech,but the traditional speech enhancement algorithm has limitations,which cannot meet the needs of practical application.Therefore,in recent years,speech enhancement based on nonnegative matrix factorization is the focus of many scholars at domestic and abroad.The main work of this thesis is to research and analyze the speech enhancement algorithms in single channel and dual channel systems under different background noise environment,and carry out the research of nonnegative matrix factorization speech enhancement algorithm.Firstly,this thesis analyzes different background noise models,describes two kinds of speech signal processing methods,speech separation and speech enhancement,which can realize the target speech extraction,and expounds the nonnegative matrix decomposition algorithm and speech quality evaluation method under different noise models.Secondly,when the components of speech and noise are similar,the traditional non negative matrix factorization speech enhancement algorithm does not have distinguishing characteristics.In this thesis,a multi parameter regularization constraint is introduced into the cost function,which makes the decomposed dictionary retain more speech characteristics.In this thesis,a nonnegative matrix factorization speech enhancement algorithm based on single channel fully supervised model is proposed.The main idea of the algorithm is to train the pure speech and noise respectively by using the regularized constrained non negative matrix factorization algorithm to generate the dictionary matrix with the characteristics of distinguishing noise and pure speech.Then,the multi-parameter regular constrained nonnegative matrix factorization algorithm is used to enhance the noisy speech spectrum to obtain the estimated speech spectrum.Finally,the masking phase compensation algorithm is used to further improve the quality of the reconstructed target speech signal.Then,considering that the application scope of the fully supervised model is limited in the complex noise field,a dual channel non negative matrix factorization speech enhancement algorithm under the unsupervised model is proposed.In order to solve the problem that the speaker information cannot be recognized in coherent and incoherent noise environment,this thesis combines the improved speech separation algorithm based on generalized crosscorrelation nonnegative matrix factorization with the phase estimation algorithm based on posterior signal-to-noise ratio to achieve the effective separation of speech and noise,so as to achieve the effective extraction of target speech.The experimental results show that the improved generalized cross-correlation nonnegative matrix factorization speech enhancement algorithm can greatly improve the content intelligibility and speech quality of the speaker.Finally,aiming at the situation that it is difficult to get effective speaker features in complex sound field,a single channel speech enhancement method based on semi supervised model and binary masking is proposed.In this method,only noise is trained by using timefrequency dictionary.This method uses sparse non negative matrix factorization algorithm to train the background noise for many times,and then completes the preprocessing of noisy speech by binary masking of nonnegative matrix factorization,and reconstructs the target pure speech signal by binary masking,so as to realize speech enhancement.Experimental results show the effectiveness of the algorithm.