Speech denoising using Bayesian NMF with online base update

Abstract

A new speech denoising method based on online non-negative matrix factorization (NMF) is proposed in this paper. To achieve an efficient model for the temporal dependencies of speech and noise, and to improve the robustness for the actual non-stationary noisy environments, the Bayesian NMF is extended to the proposed model and a new noise basis matrix online update method is exploited. Firstly, the speech basis matrix is pre-trained off-line with the Bayesian NMF method. In speech denoising stage, the noise basis matrix is continuously updated by utilizing the noise frames in the noisy observation with the Bayesian NMF. The noise basis matrix is initialized via a pre-trained universal noise NMF model and the noise data for the matrix adaption are selected using a likelihood ratio test (LRT) speech decision criterion. Then the updated noise basis matrix and the pre-trained speech basis matrix are employed to the enhancement of the noisy signal. Finally, to address the incomplete separation and the speech distortion problem, a speech activity probability based noise suppression filter is presented to further eliminate the residue noise in the enhanced result. The experiment results show that the proposed method outperforms the comparative denoising algorithms in terms of objective measurement.