The Research Of Blind Source Separation Of Dynamic Number Of Speakers

Blind source separation (BSS) is an emerging technique of array signal processing and analysis, which aims to recover the unobserved signals or sources from the observed mixtures without any prior knowledge. It is widely used in wireless communication, medical signals processing, image enhancement, audio mixtures separation and some other domain.BSS for audio mixtures is the original intention of this technique and a challenging problem in signal processing realm. There are many different aspects about it corresponding many different algorithms. In this thesis, the principles and algorithms of blind source separation are firstly discussed and then the main work are as follows:Most approaches of blind source separation in speech suit the situation that the number of speakers is known and constant during the process of separation but it is dynamic in reality. This thesis uses only two microphones to separate speech signals of dynamic speakers. The method can not only reflect the changing number of speakers but also solve the problem of overcomplete and undercomplete caused by the different numbers of speakers and microphones. The essence of this method is that: Firstly, we estimate the number of speakers and according to the different numbers we can separate the received mixed signals to many segments. Compared with several other ways of estimating the number of speakers, we choose cross-power spectrum phase(CSP) for its rapid speed and simpleness. Secondly, we do blind source separation for each segment. Because the number of speakers is dynamic, it is not always equal to that of microphones, we use degenerate unmixing estimation technique (DUET). Finally, we unite every speaker's voice from each segment by means of speaker identification which is our proposal.Meanwhile we use this method for blind source localization .Based on CSP,...