FAST MULTICHANNEL NONNEGATIVE MATRIX FACTORIZATION WITH CONSTRAINTS ON ACTIVE SOURCE CANDIDATES

This paper presents an extension of the model of multi-channel nonnegative matrix factorization (MNMF) by introducing constraints on active source candidates in the time-frequency (TF) bins. An algorithm of the proposed MNMF considerably faster than the conventional one is then developed under the constraint that at most two sources are active in each TF bin. To reduce the number of matrix operations in MNMF, the proposed algorithm simultaneously diagonalizes each pair of two active source spatial covariance matrices. As a result, the computational time complexity of the fast MNMF is comparable to those of TF mask approaches without assuming the sparsity constraint of source spectra unlike them. The experimental results show that the fast MNMF is only about twice as slow as the corresponding TF mask approach while maintaining the separation performance of the original MNMF.