Soft decision based Laplacian model factor estimation for noisy speech enhancement

Abstract

The Laplacian model factor estimation is a critical link for noisy speech enhancement technique employing Laplacian statistical model priori of clean speech. In this letter, we propose a novel estimation algorithm for this parameter based on soft decision in discrete cosine transform domain. As the speech signal is not always present in the noisy speech signal at all components, we first compute the speech presence probability which is decided in each discrete cosine transform component, and then based on the minimum mean square error estimation theory, the Laplacian model factor is estimated in the speech presence stage. Simulation experiment results demonstrate that the proposed algorithm possesses improved performance than that of the conventional method under different noisy conditions and levels.