Eigenvalue Filtering Based Subspace Approach For Speech Enhancement

A novel eigenvalue substitution and filtering subspace based speech enhancement approach is proposed for a better reconstructed speech quality. On the basis of GEVD(the generalized eigenvalue decomposition) subspace approach for speech enhancement,we improve speech quality by eigenvalue substitution and filtering. The study shows that after the generalized eigenvalue decomposition for the covariance matrices of the clean speech signal and the noise, the larger eigenvalue component contains the main information of speech signal, while the smaller eigenvalue component contains noise information. After the eigenvalue ordering, we replace the smaller eigenvalue and its corresponding eigenvectors with the larger ones or filter the smaller eigenvalue components so that there is a better enhanced speech signal and the correlation of the speech frame can be improved greatly. Compared with other subspace methods, the proposed methods can restrain strong noise more effectively, get a higher SNR(Signal Noise Ratio) and ensure a better aural intelligibility.The innovations are described as follow:1. We propose an eigenvalue component filtering subspace approach for speech enhancement to improve speech quality by filtering small eigenvalue components, based on GEVD subspace approach.2. We propose an eigenvalue component substitution subspace approach for speech enhancement, which replaces the larger eigenvalue component with smaller one after the generalized eigenvalue decomposition for the covariance matrices of the clean speech signal and the noise.3. After eigenvalue filtering or substitution, the remainder eigenvector is no longer a square matrix. In the common case, we can not get its inverse matrix, so we introduce the generalized inverse matrix operation to solve this issue.