Perceptually motivated speech enhancement based on constrained optimization

This paper is concerned with recently proposed perceptually constrained signal subspace (PCSS) method for speech enhancement. Optimal linear estimator minimizes average speech distortion power for a given set of the perceptually motivated constraints on the residual noise power spectrum. We also derived frequency-domain version of the PCSS method. This is done using a special assumption about structure of the covariance matrices. Surprisingly, the resulting estimator is almost identical to the well known IND (Inaudible Noise Distortion) rule. In other words we found the conditions for which an empirical IND rule is optimal. The experiments show that a intelligibility of the enhanced speech is similar for both methods, however in the case of noise attenuation the subspace-based methods are superior.