Noise Tracking by Exploiting DFT-Domain Subspace Decompositions

DFT domain based noise reduction algorithms can be effective for noise reduction in various speech processing applications. In general these algorithms apply a noise-PSD dependent estimator to the noisy speech DFT coefficients in order to estimate the clean speech DFT coefficients. Since the noise PSD is unknown in advance, estimation is one of the crucial elements of such a noise reduction system. In this paper we discuss a method for estimation of the noise PSD while both speech and noise are present. This method is based on DFT domain based subspace decompositions of noisy correlation matrices that are estimated per frequency bin in the DFT domain. The proposed algorithm shows increased tracking performance of the noise PSD in comparison to minimum statistics. Further, listening tests show a preference for the proposed method over MS.