DCT-Based Speech Enhancement Using a New Speech Variance Estimation

This paper proposes a new speech variance estimation algorithm for speech enhancement which operates in the discrete cosine transform (DCT) domain. The estimation method is based on the definition of variance and derived under the simple Gaussian distribution assumption for noise and speech components, which is in agreement with the DCT-based speech enhancement framework. The proposed method can be updated adaptively without making use of a weighting factor; it overcomes the disadvantage in the original algorithm that the trade off between the noise reduction and speech distortion brought into the signal can not be satisfied availably. A number of subjective and objective tests under various conditions show the improved performance of the proposed algorithm