Two-stages Predictive And Separation Method For Underdetermined Blind Source Separation

Blind Source Separation(BSS)was originally derived from cocktail party problem.It has developed rapidly within the past 30 years due to its wide application.According to the relationship between the number of source signals and observed signals,the problem of BSS can be divided into Overdetermined Blind Source Separation,Standard Blind Source Separation and Underdetermined Blind Source Separation(UBSS).In real world,the number of sound acquisition devices is generally less than the number of speakers.Therefore,the study about UBSS is more meaningful and challenging.The K-C-means and adaptive K-C-means clustering algorithms are introduced in this paper.Because of our focus on the difference orientations of sample points,the cosine distance is used to measure the similarity of the sample points.We call it K-C-means clustering algorithm.The adaptive K-C-means clustering algorithm finds the optimal number of clustering based on the relationship between the total clustering error and the number of clustering centers.Short-Time Fourier Transform(STFT)is a time-frequency analysis method.By windowing the time-domain signal and applying the Fourier Transform,the time and frequency information of the signals can be obtained.There are mainly the following three aspects:Part ?: A Time-Frequency Two-Stages method(TFTSM)is proposed for linear mixed UBSS problem with non-sparse signals.Since the number of source signals is unknown in advance,the first step is to estimate the number of source signals and the mixed matrix.A Maximum and Minimum Value algorithm is presented to filter out the Simple Signal Points of signals in the time-frequency domain.Adaptive K-C-means clustering algorithm is used to determined the number of source signals and estimated the mixed matrix.Numerical experiments show that the mixed matrix can be found with high accuracy.Part ?: In the second step of the TFTSM,the improved Plane Pursuit algorithm in Time-Frequency domain is proposed in this paper.At each Time-Frequency point,the hyper plane that best matches the observed signals is found.Then it is used to recover source signals.We also put forward the error correction to solve the situation that the number of source signals changes or the number of source signals is estimated incorrectly.Numerical experiments show that the algorithm can effectively recover the source signals.Part ?: The idea of BSS is applied to speech enhancement.Each type of noise is regarded as a kind of source signal.TFTSM is used to separate the observed signals.Then the target signals without noise sources are recovered,so as to achieve speech enhancement.In the first chapter,we introduce the Adaptive K-C-means clustering algorithm and Short-Time Fourier Transform which are important in our method.In the second chapter,we explain how to accurately estimate the number of source signals and the mixed matrix through TFTSM,then recover the source signals.The third chapter describes how to achieve speech enhancement through our proposed algorithm.The noise signals are separated from the target signals as different kinds of special source signals.Numerical experiments and analysis illustrate the effectiveness and accuracy of our algorithm,and finally gives our conclusions and prospect of the future work.