Research On The Method Of Underdetermined Blind Speech Separation Based On NMF And Sparseness

The blind source separation(BSS)technique is a process of separating the source signal from the observed signal in cases where the prior information is limited,and the original signal,mixing type,channel information,etc.are unknown.Blind source separation algorithm is usually used to solve the determined or over-determined problems,i.e.,the number of the observed signals and the original sources is equal or the former is more than the latter.However,in practice,it tends to be under-determined problems,i.e.,the latter case happens more often than the former one,so it is very important and valuable to study the algorithm of the under-determined blind source separation algorithm.Among them,the original purpose of the non-negative matrix factorization(NMF)algorithm is used to solve the problem of under-determined blind source separation,which does not depend on the characteristics of independence or non-Gaussian of the signals.Speech as one of the objects of blind source separation,in the computer auditory,video conferencing,information anti-interference and bio-medical research,and many other areas,has shown a broad development prospects.Therefore,this thesis explores and improves the NMF algorithm based on the problem of under-determined blind source separation in the linear hybrid mode.The core contents consist of three aspects below:(1)Based on the model of single-channel blind speech separation,the basic NMF algorithm with the objective function based on Euclidean distance or Kullback-Leibler(KL)divergence,and the optimization NMF algorithm of adding Short-Time Fourier transform(STFT),which is called as the SNMF algorithm,are given.Then the different objective functions are selected and simulated.Compared with the NMF algorithm of Euclidean distance and its improved algorithm,the NMF algorithm of KL dispersion gives better separation.(2)For the model of under-determined instantaneous mixing of speech signals,a blind speech separation algorithm based on KL-SNMF with additive constraints is given.Firstly,being use of its unique short-term stability,do pre-emphasis,sub-frame and windowing,and some other handles.Then,the amplitude spectrum is taken as the input matrix to reduce the dimension.Finally,using the negative entropy as the objective function,and the fast fixed point independent component analysis(Fast ICA)algorithm which is optimized through Newton iteration method,do speech separation.(3)For the model of under-determined convolution mixing of speech signals,a blind speech separation algorithm based on EM-KL-SCNMF is given.Firstly,the convolution non-negative matrix factorization(CNMF)algorithm is applied to the blind speech separation model,which retains the speech features of information and the correlation between frames better.Then,in order to make full use of the redundancy between channels,and get rid of the constraints of source with statistical independence and non-Gaussian distributions,the Expectation-maximization(EM)algorithm is employed for iterative optimization,to minimize the objective function and further to get the estimated signal;At last,using Inverse STFT to produce the separation filter,together with reconstructing signals,get the time domain separation speeches.In this thesis,we mainly study the three aspects above.In comparison with the separation of a single signal and its original signal,using a large number of comparative simulation experiments to prove the feasibility and effectiveness of the proposed algorithms,which can keep the signal waveform better and has a large and stable correlation coefficient,which can realize the signal separation better.