Subspace speech enhancement based on minimum statistics noise estimation

A novel method for estimating noise on eigenvectors for subspace speech enhancement is proposed. The method is unlike with other methods and does not require explicit voice activity detection and estimation of the noise covariance matrix. The method can track eigenvalue minima on each eigenvector without any distinction between the speech activity and the speech pause, thus updating the noise estimate throughout the entire signal. This allows a more accurate noise estimate to be produced and improves the quality of the enhanced speech. Simulation results show that the proposed method is excellent than the subspace methods in terms of log-spectral distortion (LSD), perceptual evaluation of speech quality (PESQ) and informal listening tests.