| This thesis mainly deals with the problem of blind deconvolution with only second-order statistics in the frequency domain. Mixture signals in the real world are always the convolution of different sources. The convolutioe mixing in the time domain can be converted into instantaneous mixing in the frequency domain. Algorithms for separating instantaneous mixtures can then be applied in each frequency bin. The main problem of this approach is that the separated outputs in each bin are in a different order. This permutation ambiguity prevents reconstruction of signals by inverse Fourier transform. In order to overcome this problem, the learning procedure of our adaptive algorithm is processed in the time domain, while all the others are in the frequency domain. In the simulations of blind deconvolution and blind echo cancellation, this algorithm has better performance than those where the learning procedure is in the frequency domain. |
