Speech Enhancement Based on Laplacian Speech Modeling and Unscented Kalman Filtering

Considering nonlinear and non-Gaussian speech signals, a novel speech enhancement algorithm is proposed based on Laplacian speech modeling and unscented Kalman filtering (UKF) using a single microphone. The probability density function (PDF) of speech signals in the time domain is better modeled by a Laplacian rather than a Gaussian density by analyzing statistical property for clean speech signals. So in this paper the PDF of clean speech signals is assumed to Laplacian density and the PDF of noise are modeled by Gaussian distributions. Speech enhancement algorithm based on UKF is used to solve non-linear property for speech signals. Simulation results demonstrate that the proposed algorithm possesses good performance both in objective and subjective tests with colored noise and white noise presenting.