A Novel Frequency Domain Linearly Constrained Minimum Variance Filter for Speech Enhancement

A reliable speech enhancement method is important for speech applications as a pre-processing step to improve their overall performance. In this paper, we propose a novel frequency domain method for single channel speech enhancement. Conventional frequency domain methods usually neglect the correlation between neighboring time-frequency components of the signals. In the proposed method, we take this correlation into account via: 1) considering neighboring correlation for speech signals, we break down the clean speech into two uncorrelated components; 2) considering neighboring correlation for noise, we approximate the noise as a rank-1 component. Then, we design a linearly constrained minimum variance (LCMV) filter which aims at removing the dominant part of the noise, while keeping the speech signal undistorted. Performance of the proposed method is evaluated in terms of output signal to noise ratio (SNR) and speech distortion index under various noise environments. Evaluation results demonstrate that our method yields higher noise reduction and lower speech distortion compared to some recent methods.

Keywords:

Neighboring Correlation , Linearly Constrained Minimum Variance Filter , Reduced Rank Model , Speech enhancement