Subspace-Based Speech Enhancement using Triangular Matrix Decomposition and Noise Variance Estimation

In this paper, the speech enhancement technique based on signal subspace approach is presented. Decomposition of the noisy speech vector space into a signal-and noise subspace is usually performed using singular value (or eigenvalue) decomposition (SVD or EVD). Rank-revealing ULV decomposition is used in the paper as computationally attractive alternative to the SVD. Minimum variance (MV) estimator of the signal matrix using ULV decomposition is considered. In comparison with the original technique the noise estimation approach is improved and the refinement step is renewed. Simulations results show that the proposed ULV-based approach provides the higher performance than the previous one.