Adaptive Sparsity Non-Negative Matrix Factorization for Single-Channel Source Separation

A novel method for adaptive sparsity non-negative matrix factorization is proposed. The proposed factorization decomposes an information-bearing matrix into two-dimensional convolution of factor matrices that represent the spectral dictionary and temporal codes. We derive a variational Bayesian approach to compute the sparsity parameters for optimizing the matrix factorization. The method is demonstrated on separating audio mixtures recorded from a single channel. In addition, we have proven that the extraction of the spectral dictionary and temporal codes is significantly more efficient with adaptive sparsity which subsequently leads to better source separation performance. Experimental tests and comparisons with other sparse factorization methods have been conducted to verify the efficacy of the proposed method.