Single-channel Speech Separation Using Dictionary-updated Orthogonal Matching Pursuit and Temporal Structure Information.

In this paper, we propose a two-stage sparse decomposition-based method for single-channel speech separation in time domain. First, we propose a Dictionary-updated orthogonal matching pursuit (DUOMP) algorithm which is used in both separation stages. In the proposed DUOMP algorithm, all atoms of each source-specific dictionary are updated by subtracting off the current approximation of each source to the original atoms. It is proved that the DUOMP algorithm can limit the separated sources within a region where they are uncorrelated in statistical sense more quickly. Then, we propose an adaptive dictionary generation method followed by a frame labeling method to perform a second-stage separation on the mixed frames having certain temporal structure. Experiments show that the proposed method outperforms a separation method using sparse non-negative matrix factorization (SNMF), a separation method using OMP and a source-filter-based method using pitch information in overall. Additionally, what affects the performance of the proposed method is also shown. [ABSTRACT FROM AUTHOR]