Sparseness constraints on nonnegative tensor decomposition

The Nonnegative Tensor Factorization (NTF) method, has been shown to separate the mixture of several sound sources reasonably well. Based on the sparsity of power spectrogram of signals, we propose to add sparseness constraints to one factor matrix, which contains frequency basis, to obtain a sparse representation of this nonnegative factor. Two methods are presented for the sparse and nonnegative factor matrix. The algorithms are tested to separate the instrument sources from a two channel mixtures containing a clarinet, a piano and a steel drum. Some comparison results of these two NTF-based methods are then presented.