The speech in real environments is usually interfered by the noise. Although microphone array speech enhancement algorithms have been much studied to reduce the noise and enhance the speech quality, they can not completely meet all the application requirements. The main work of this thesis is to do the research about two microphone array speech enhancement algorithms: the post-filter algorithm and the signal subspace algorithm. The key of the post-filter algorithm is the estimation of the post-filter and the core of the signal subspace algorithm includes the subspace selection and linear filter estimation. This thesis solves these key problems based on the masking properties of human ears. The main contributions and novelties include: The noise on the microphones is divided into two categories: correlated noise and uncorrelated noise. We analyze the characters of the auto- and cross-power spectrum of the received signals on the array and give a novel post-filter. Then, the post-filter is expressed in matrix style and the masking properties of human ears are incorporated to improve the post-filter. It results in the further improvement of the post-filter performance. The conventional thresholds based subspace selection method is not accurate. Based on the characters that the eigenvalues in noise subspace should be equal, we propose a better subspace selection method. The noise power spectrum is estimated by the conditional probability in noise subspace. We use the masking properties of human ears to estimate the values of Lagrange multipliers and a novel linear filter is proposed. We use the Gaussian, Laplacian and Gamma models to describe the distributions of the speech and noise. A new subspace selection method is proposed by maximizing the existence probability of the speech. The noise power spectrum is estimated according to the speech existence probability. The masking properties of human ears are used to balance the speech distortion and the residual noise and a novel post-filter is proposed. The research of microphone array speech enhancement algorithms based on the masking properties of human ears is the highlight of this thesis. The proposed algorithms in this thesis are much better than the conventional algorithms on both aspects of speech distortion reduction and noise reduction. The expected targets are achieved.
修改评论