Nonnegative Matrix Factorization (NMF) has become an increasingly popular method in the field of non-stationary speech denoising. However most NMF based algorithms assume prior knowledge about the background noise, which is often not available in the time-varying and mobile environments. In this paper, we propose a two-step NMF based speech-noise separation algorithm to address this issue. This algorithm takes the outcome of the first NMF separation as the dataset to train the basis vectors for the background noise, which will be used for the second-step NMF separation with fixed speech and noise basis vectors. Experimental results show that the proposed algorithm could achieve better results than other NMF algorithms for speech-noise separation.
Index Terms: single channel speech separation, nonnegative matrix factorization, voiced/unvoiced sound classification, Wiener filter.