Speech enhancement based on speech spectral complex Gaussian Mixture Model

This paper presents a speech enhancement approach based on speech spectral complex Gaussian Mixture Model (GMM). First, a construction algorithm of speech spectral GMM is introduced and it is based on the distance measure of speech spectral Gaussian probability. Then a noise estimation algorithm based on the GMM is proposed in the Maximum Likelihood criterion using the Expectation-Maximum (EM) algorithm. Speech enhancement experimental results show that the GMM-based MMSE estimators, especially the GMM-based MMSE short-time spectral estimator, can afford better performance than alternative speech enhancement algorithms and the proposed noise estimation algorithm can improve the enhancement performance more, especially at low SNRs.