An eigenvalue filtering based subspace approach for speech enhancement

In this paper, a subspace approach based on eigenvalue filtering is proposed for enhancement of corrupted speech. The new method firstly simultaneously diagonalizes the covariance matrix of clean speech and noise signal based on GEVD (generalized eigenvalues decomposition), and then filters the smaller components whose eigenvalues are less than zero. Because the remainder eigenvector matrix after filtering is irreversible, we introduce the generalized inverse matrix transform to solve this problem for recovery of speech signal. Experimental results show the proposed method performs better than many conventional methods under strong noise conditions, in terms of yielding less residual noise and lower speech distortion.