Signal subspace approach for psychoacoustically motivated speech enhancement

In this paper we deal with the perceptually motivated signal subspace methods for speech enhancement. We focus on extended spectral-domain-constrained (SDC) estimator. It is obtained using Lagrange multipliers method. We present an algorithm for a precise computation of the Lagrange multipliers, allowing for a direct shaping the residual noise power spectrum. In addition the SDC estimator is presented in a new, possibly more effective form. As a practical implementation of the estimator we propose perceptually constrained signal subspace (PCSS) method for speech enhancement. The approach utilizes masking phenomena for residual noise shaping and is optimal for the case of coloured noise. Also, less demanding approximate version of this method is derived. Finally comparative evaluation of the most known subspace-based methods is performed using objective speech quality measures and listening tests. Results show that the PCSS method outperforms other methods providing high noise attenuation and better speech quality.

Research highlights
► Extended spectral-domain-constrained estimator for speech enchancement is considered.
► Lagrange multipliers are directly computed allowing precise residual noise shaping.
► The residual noise is shaped optimally according to the masking phenomenon.
► A novel perceptually constrained signal-aubspace (PCSS) method is proposed.
► Finally comparative evaluation of the selected subspace-based methods is performed.