Improved expectation-maximization framework for speech enhancement based on iterative noise estimation

Recently, our team developed a novel Expectation Maximization (EM) framework for speech enhancement. It gives a significantly improved estimation of the speech power spectrum that outperforms many traditional approaches. In this paper, we further extend the EM framework by including an efficient iterative noise estimation algorithm, which improves the estimation of the noise power spectrum from the noisy observation. Besides, we notice that some speech frames, particularly those with high signal to noise ratio (SNR), need to be monitored closely during the iterative enhancement process, or spectral distortion may result. A stopping criterion is thus developed to stop the iteration when a good result has been achieved. Experimental results show that the new approach gives a significant improvement over the original EM framework and also traditional speech enhancement methods.