A greedy approach to Linear Prediction with sparse residuals

This paper focuses on the problem of Linear Prediction (LP) constrained by sparse residuals. After reformulating the problem to finding the largest linear correlated strict subset in a given vector set, a greedy method is proposed to determine the support of the sparse residuals iteratively by testing each entry with respective temporary prediction error. The greedy method is then simplified to reduce computational cost. Compared with reference algorithms and conventional LP model, the proposed methods are tested in the speech coding scenario. Experiment results demonstrate that the proposed greedy methods work well and suggest that LP with sparse residuals provides accurate estimation, and is much practical in the scenarios that more bits are allocated for coding residuals.