Study On Frequency Domain Blind Source Separation Based Speech Enhancement Methods

In our lives, speech signals are often disturbed by various interferences such as background noise and room reverberation. The existing of noise and reverberation not only affects human hearing, but also has influence on other steps of speech signal processing. So it is important to enhance speech using signal processing technology. In fact, besides speech enhancement, speech separation can also get rid of the influence of noise and reverberation. Blind Source Separation (BSS) is a challenging subject and becomes to be a popular research area in signal processing field in recent years. In the last two decades, a number of algorithms addressing the instantaneous BSS problems have been proposed and gained some results. However, current results are far from its mature solution. Moreover research on real-world convolutive speech signals has just begun. This thesis focuses on the methods of frequency domain blind source separation based speech enhancement which include three parts:The Independent Component Analysis (ICA) -based BSS approach seems to be a very flexible and effective technique for source separation, but it has an inherent disadvantage in that there is difficulty with the slow convergence of nonlinear optimization. In this thesis, a new algorithm is introduced for blind source separation, in which ICA and beamforming are combined to resolve the slow-convergence problem through optimization in ICA. The temporal utilization of null beamforming through ICA iterations can realize fast- and high-convergence optimization. The results of the signal separation experiments reveal that the signal separation performance of the proposed algorithm is superior to that of the conventional ICA-based BSS method, even under reverberant conditions.Blind separation of human speech with music signals is a difficult problem. The separation of ICA deteriorates for real-world convolutive mixtures. Hence, the separated signals can have cross-talk components. An algorithm for BSS of convolutive mixtures is introduced here. Separation of signals is performed in two stages. The first stage involves the application of an ICA algorithm, and shrinkage functions are applied to a set of wavelet coefficients in the second stage. The results show that the technique can separate music from speech signals with highly acceptable results.A speech enhancement method based on independent component analysis with reference(ICA-R) in subband and shrinkage function post-processing is proposed, which is used innoise and reverberation environments. Combining ICA-R and shrinkage function, two channel noisy signals were processed to obtain the enhanced target signal. Firstly, two channel noisy signals were passed through the analysis filter banks to generate the subband noisy signals. Then, ICA-R was used in subband to extract the estimated target signal we were interested in. Finally, after synthesis filter banks, shrinkage function was used to enhance estimated target signal further. The experimental results show that the proposed method is effective and the distortion of the target signal is small.