Perceptually constrained signal subspace method for speech enhancement - approximate solutions

摘要

We presented the perceptually constrained signal subspace (PCSS) approach for speech enhancement. As the exact esti- mator is computationally demanding we also proposed sub-optimal realizations of the PCSS method. The first approach replaces the covariance matrix of the noise energies in the transform domain with a diagonal matrix. Unlike the exact method the approximate solution does not require whitening, thus the number of operations per frame can be significantly reduced. The experiments show that degradation due to approximation depends on noise type and can be neglected for white-like noises. We also derived frequency-domain version of the PCSS method using circulant covariance matrices. Resulting estimator seems to be almost identical to the well known the IND rule. The experiments show that perceptually constrained methods offer similar speech intelligibility for different noise environments and perform significantly better than conventional non-perceptual method. However the signal subspace-ba-sed approaches outperform the frequency domain method in terms of noise attenuation.