The Study On Underdetermined Blind Source Separation Of Mixed Speech Separation

In real world, speech may be mixed by noise or other speech. Speech enhancement technology can separate speech from the mixtures. Blind source separation (BSS) is an important speech enhancement technology. It separate signals from mixtures without knowing the mixing system. And it is widely used in communication, biomedical engineering and speech processing. The famous algorithm of blind source separation is Independent Component Analysis method (ICA). However, ICA is a complete problem where the number of observed signals is no less than sources. When the number of observed signals is less than sources, the blind source separation problem is called underdetermined blind source separation. The underdetermined case is more common in the real world.In this paper, underdetermined blind source separation technology is researched. This paper including three parts as follows:(1) A mixing matrix estimation method based on single signal detection. The problem of mixing matrix estimation is the first step of underdetermined blind source separation. In this paper the mixing matrix estimation methods such as potential function method and K means clustering method are studied. By analysis their faults, a method based on single signal detection method is proposed. In this method, the samples after single signal detecting are clustered. The result of this method is more accurate than K means cluster.(2) A separation method based on the sparsity of signals. The most important sparsity based method is minimum l1 norm method. In the samples when the number of active sources is more than the number of observed signals the minimum l1 norm method can not work accurately. In this paper, these samples are shielded first and then the left samples are separated by the minimum l1 norm method. The compute experimental results show that this method can separate signal with high quality.(3) An underdetermined blind separation method based on relaxed sparsity of signals. Clearly, in the sparsity separation method, the qualities of separated signals depend on the sparsity of signals. So when the signals are non-sparse, these methods can not work well. In this paper, the method based on subspace is proposed. This method relaxed the condition of sparsity of signals. The experiments show that this method is effective.