Approach based on colored character to blind deconvolution for speech signals

Abstract

The time-based convolutive mixing can be transformed to the frequency-based instantaneous mixing in blind problems as the convolution operation in time domain corresponds to a multiplication in frequency one. After performing blocks on the whole speech signal, and applying Fast Fourier Transform (FFT) on each blocked signal, the new colored instantaneous mixing signals in frequency domain are constructed. Making use of the Second Order Statistics (SOS), the unmixed matrices are respectively computed at each frequency bin by blind separation based on cost function which is aimed at colored signals. Thus, in time domain, deconvolution is performed after employing correlation-based permutation to the inferred signals which are associated with the unmixed matrices at different frequency bins. The experiment proves the effectiveness of the approach to blind deconvolution by blind separation in frequency domain.