A signal subspace approach for speech enhancement

Abstract

A perceptually based linear signal estimator for enhancing speech signals degraded by uncorrelated additive noise is developed. The estimator is designed by minimizing the signal distortion while maintaining the residual noise level below some given threshold. The estimator is shown to be a Wiener filter with adjustable input noise level. This level is determined by the threshold of the permissible residual noise. The estimator is implemented using the signal subspace approach. The vector space of the noisy signal is decomposed into a "signal subspace" and a complementary orthogonal "noise subspace." Estimation is performed from vectors in the signal subspace only, since the orthogonal subspace does not contain signal information. The proposed estimator is shown to be a refinement of a version of the spectral subtraction signal estimator. The latter estimator is shown to be arymptotically optimal for stationary signal and noise in the linear minimum mean square error sense.