A speedup method for the separation of speech signals in frequency domain

The independent component analysis (ICA) is a commonly used method to find the demixing matrix for the blind source separation (BSS). For speech signals, we should solve BSS problems in the convolutive mixing model, i.e., ICA technique is extended to the frequency domain. The cross-spectral density matrices are computed for each frequency bin instead of covariance matrices in time domain. The joint approximate diagonalization (JADIAG) algorithm proposed by D. T. Pham has been proved to be effective in dealing with the convolutive mixing problem. This paper presents a method to speed up the JADIAG computation in two phases. First, the critical band property of human auditory system is applied so that a set of selected demixing matrices is shared in a critical band to reduce the number of demixing matrices. Second, an efficient estimation of transformation matrix is proposed so that the iterations for finding the demixing matrices in JADIAG algorithm are reduced. The experiment shows that about 71% of computation time can be reduced.