Perceptually Constrained Subspace Method for Enhancing Speech Degraded by Colored Noise

In this paper we present a novel method for enhancing speech corrupted by colored noise. A recent extension of signal subspace approach to colored-noise processes is employed. Enhancement is performed using optimal linear estimator which minimizes average signal distortion power for a given set of constraints on the residual noise power spectrum. Perceptual criteria give lower speech distortion than SNR-based solutions. Thus, our proposition is to use constraints defined in DFT domain that are consistent with masking properties of the human ear. Optimal filter is found by solving the constraints' equations for given masking threshold. Proposed method utilizes currently most advanced ideas in signal subspace speech enhancement and is optimal in the general case of colored-noise process.