Research On Two-channel Speech Enhancement Method Based On Non-negative Matrix Factorization

In the harsh acoustic environment,due to the influence of noise,the obtained speech signal will be filled with impurities,and the quality and speech intelligibility will be seriously affected.Therefore,speech enhancement is required.In the previous speech enhancement algorithms,it is possible to remove the voice signal part while removing the noise,affecting the performance of the speech processing system.At present,signal processing is widely used in the field of scientific engineering,especially the research on speech signals has become a hot topic in recent years.As a major branch of speech signal processing,the goal of speech enhancement algorithms is to improve speech quality and speech intelligibility caused by signal interference due to various reasons.Based on the development of traditional speech enhancement algorithms,several new algorithms have been proposed in conjunction with machine learning in recent years.Non-negative matrix factorization(NMF)has become one of the more widely used of these algorithms.In addition to machine learning,microphone array signal processing techniques are also widely used in speech enhancement algorithms,such as adaptive beamforming.In view of this,this thesis combines the non-negative matrix factorization and the related algorithms of the microphone array,uses the idea of blind source separation to enhance the speech signal,and forms a two-channel speech signal enhancement method based on the non-negative matrix factorization,it can post-process the sound signal.First,the generalized cross-correlation method of the microphone array is used to obtain the estimated DOA.Then,the non-negative matrix method is improved to improve the initialization of non-negative matrix factorization.The traditional non-negative matrix factorization generally uses random initialization,but the frequency domain structural characteristics of the speech signal cannot be maintained in the application of the speech enhancement algorithm,and the initialization without guidelines under a limited number of iterations will cause alarge loss of the final result.This uses the recursive Kmeans++ algorithm to initialize the NMF,and instead of using the traditional Euclidean distance during the initialization process,it replaces it with another new distance formula.The coefficient matrix of the non-negative matrix factorization is processed by the estimated arrival delay,and a cubic generalized cross-correlation method is proposed to better estimate estimated time of arrival based on the traditional generalized cross-correlation method.Finally,the proposed algorithm is verified one by one with PyCharm.Through research and analysis and a large number of tests and experiments,it is shown that the speech enhancement algorithm proposed in this thesis has a higher signal-to-noise ratio and intelligibility than traditional methods.