A method and apparatus is disclosed for performing blind source separation using frequency-domain normalized multichannel blind deconvolution. The multichannel mixed signals are formed as frames of N samples, which consist of r consecutive blocks of M samples. The frames of mixed signals are separated using separating filters in the frequency domain in an overlap-save manner using a discrete Fourier transform(DFT). The separated signals are then converted back into the time domain using the inverse DFT to be applied to a nonlinear function. The cross-power spectra between separated signals and nonlinear-transformed signals are computed and are normalized by the power spectra of separated signals and the power spectra of nonlinear-transformed signals to have flat spectra. The invention then applies the time domain constraint to preserve the first L cross-correlations. These alias-free normalized cross-power spectra are further constrained by nonholonomic constraints. The invention then computes natural gradient by convolving alias-free normalized cross-power spectra with separating filters. After the length of separating filters is constrained to L, separating filters are updated using the natural gradient and normalized to have unit norm. The terminating conditions are checked to see if separating filters converged. |