Blind Speech Signal Separation Technology Research And Application

Blind speech separation is an approach, which only using information about the mixtures observed signals of every input channel for estimating source signals. The estimation is performed without possessing information on each source, or on how the sources are mixed, mainly for communication area. As human accelerated pace into the information society, blind speech separation is used in more and more areas. Blind speech separation studies are being launched in the time domain and frequency domain.The speech signals under an instantaneous mixing model can be separated in the domain.But when they under an convolutive mixing model, separating in the frequency domain will perform better. Now, there are many blind source separation algorithms,mainly divided into two categories, batch algorithm and adaptive algorithm. Joint approximate diagonalization of eigenmatrices(JADE) is an important algorithm in batch algorithm,and adaptive algorithm has gradient algorithm. Many scholars working on these two categories of algorithm improvements, so Fast ICA algorithm became, which combines the advantages of the batch algorithm and adaptive algorithm. Like with other ICA algorithms, the permutation and scaling ambiguities of the Fast ICA algorithm should be aligned so that the separated signals are constructed properly in the time domain. For this part of the difficulty, this article do lots of simulation and comparative study. Finally,achieve the alignment of the frequency and eliminate the uncertainty of the amplitude.In this thesis, following several studies has been conducted:1. Analysis and compare four independence criterion of ICA algorithm, as maximum likelihood(ML) criterion, maximum information(infomax) criterion,minimum mutual information criterion and maximum non-Gaussianity criterion.Analysis batch algorithm and adaptive algorithm.2. In-depth study the Fast ICA algorithm. Coefficient vector orthogonality is used to guarantee uniqueness of the extracted components. Experiments result that the algorithm enjoys a number of useful properties, including fast convergence,guaranteed global convergence for certain mixing conditions and contrasts,and robust behavior when noise is present.3. For eliminating the permutation and scaling ambiguities of the Fast ICA algorithm, lots of simulation and comparative study have been done in this thesis.Finally, a simple clustering algorithm with centroids works well for grouping separated signals.4. Probing the application of blind speech separation, as in microphones. We consider the situation that the sources outnumber the microphones, or the microphones outnumber the sources.