Research On Underdetermined Blind Speech Separation Based On Sparsity In Time-frequency Domain

Currently, blind speech separation has been developed with the development of technology. We call it overdetermined or determined blind speech separation when the number of observed signals is greater than or equal with the number of source signals. Otherwise, we call it underdetermined blind speech separation. We call it linear, convolutive and nonlinear blind speech separation according to the mixed mode of the source signals. This article is most studied on linear and underdetermined blind speech separation, the main works are as follows:1、We expound the basic theory of speech signal processing and introduce the ‘two-step’ consisted of estimating the mixing matrix and recovering the source signals which uses the sparsity of speech signal in time-frequency domain. Detailed introduction about the algorithm of estimating the mixing matrix and the algorithm of recovering the source signals is done, at the same time we analyze the advantages and disadvantages of each algorithm. In addition, we introduce the algorithm of ICA which estimates the mixing matrix together with recovering the source signals.2、For the problem of the estimation accuracy of the mixing matrix, we study on the problem of speech signal’s sparsity increasing. We propose a new effective algorithm of selecting the one-winner TF points by studying on the current algorithms. At the same time, we simulate the new algorithm and prove the effectiveness of the new algorithm.3、Estimate the mixing matrix through the one-winner TF points and prove the accuracy’s improvement by simulating. We propose a new algorithm that ‘the shortest path’ combining with the binary frequency masking method for the problem of the slow speed and the low accuracy of ‘the shortest path’ in the signals’ recovery step. Finally, simulation proves the effectiveness of the new algorithm.4、We propose a new algorithm that ICA combining with the binary frequency masking method and use the speech signals’ sparsity to solve the underdetermined blind speech separation against the overdetermined and determined one by studying on the ICA algorithm. Cepstrum smooth technique has been applied to decrease the musical noise in the separated signals. At last, simulation proves that the new algorithm can solve the linear underdetermined blind speech separation and is better than DUET algorithm.