Soft Decision Based Gaussian-Laplacian Combination Model for Noisy Speech Enhancement

One of the key issues of noisy speech enhancement technique is to achieve appropriate statistical distributions to model the clean speech and noise signals accurately. Most of the existing algorithms try to employ a sole model assumption in transform domain, which, however, has been proven to being contrary with the fact. To address this problem, the statistical properties of clean speech as well as several noise signals are analyzed using actual data in Discrete cosine transform (DCT) domain, and the study indicates the statistic of clean speech DCT coefficients tending to fall somewhere in between the Gaussian and Laplacian distribution. Based on the results, a novel speech enhancement algorithm is proposed using Gaussian-Laplacian combination model, whose core is employing a linear combination of Gaussian and Laplacian distribution to model the statistic of clean speech DCT coefficients. The corresponding weights of either distribution to the combination model are adaptively adjusted in terms of the probability of each hypothesis, which is estimated based on a soft decision technique by using Bayesian theorem. Through a number of objective and subjective tests, we compare the performance of the proposed algorithm with other recent model based approaches and have found that our algorithm is superior to the related approaches at all testing environments.