Research On Underdetermined Blind Speech Separation Based On Sparse Component Analysis

With the rapid development of science and technology,the demand for speech signal processing technology is becoming more and more extensive.Many new products put much emphasis on signal processing algorithm.Thus making various algorithms of Blind speech signal processing more and more mature.Blind speech signals can be divided into linear,convolutive and non-linear way.Linear blind speech signal processing is the simplest way,but some mature algorithms can be extended to convolutive blind speech signal processing.According to the relationship between the number of receivers and the number of sources,blind speech separation can also be divided into overdetermined,undetermined,normal blind speech separation.In this paper,we put attention on linear and underdtermined blind speech separation and use sparse component analysis algorithm to divide signals.The two-step algorithm is the focus of this paper.The main works are as follows:(1)This thesis start from the basic concept of blind speech separation,analyzing specialty of speech signal on Time Domain and Frequency domain by simulation graphics.The condition when can the sparse component analysis algorithm be used will be shown in this paper.The Sparse characterization of speech signal will be analyzed.This thesis also introduce Short-time Fourier transform carefully.(2)The “two step”algorithm will be researched.The common algorithm on estimating mixed matrix and recovering the signal will be shown in this thesis.We also summed up these algorithms.(3)One of the most important work of this thesis is pretreatment on Sparseness before divide speech signals.And a new way to filter Single source point will be used.The old way to filter Single source point has been filtered.Both of the algorithms are verified by simulation.(4)The mixed matrix of single source points which are filtered by Potential function algorithm will be estimated.This thesis does experiment on signals which are mixed by three speech signals and four speech signals.(5)Shortest path,time-frequency masking algorithm will be used to recovery speech signal,analyzing the similarity between reconstruction process on compressed perception theory and blind speech separation.Based on this,we use matching tracking algorithm to recover the speech signal.All of the algorithms are verified by simulation.The signal to noise ratio and the time they spent will be compared to analysis the differences between these algorithms.We get a improved algorithm.