Research On Blind Speech Separation Based On Human Auditory System

With the development of information technology and com puter technology, there is a higher and higher demand for the efficient signal processing. In many applications, only the m ixing of signals consist of noise s and useful inform ation and could be obtained by the sensors. It’s an important problem to separate the original signal from the mixing signals. It’ s just the reason fo r the blind source separation technique produced. Currently, it has been widely used in im age processing, speech sign al processing, communication system, biomedical technology, information retrieval and other fields.We call it blind speech s eparation when we use blind source separation in speech signal processing programs. The technique of blind speech separation is to recover the original signals from the observed m ixed signals without enough knowledge about sources or transmission channel. This method often divided into overdeterm ined and underdetermined blind speech separation ba sed on the number of sources and the number of microphones. This article centered on the binaural speech separation which is one of the underdeterm ined blind speech separation methods. The main studies are as follows:(1) The basic theory of the human auditory system and speech signal processing is presented in this paper.(2) Study and analysis the theo ry of the blind speech sep aration especially the binaural blind speech separation.(3) Study,analysis and simulate the BLUES algorithm which is a classical method in binaural blind speech separation.(4) Study,analysis and simulate the algorithms for underdetermined blind speech separation. The ’two-step’ consists of estimating the mixing matrix and recovering the original signals.(5) Study, analysis and simulate the ‘two-step’ method for only two m icrophones collecting the mixing signals. For selecting the one-winner TF points, lots of work has been done and finally g et a simple and faster and more accurate method. Show the simulation results of the binary frequenc y masking method and ‘the shortest path’ method and the m ethod based on the degree of membership respectively. And finally get an improved method of recovering original signals.