A Signal-adaptive Local Cosine Transform for Source Separation by Time-frequency Masking

Time-frequency masking is often used for source separation of underdetermined audio mixtures. It depends on the fact that the sources can be represented disjointly in some transform domain. The focus of this paper is on demixing sources from instantaneous, two-channel mixtures by binary masking. We investigate trees of local cosine bases from which a suitable transform may be generated—the best basis is chosen by a computationally efficient algorithm and is adaptively selected to match the time-varying characterists of the signal. Our heuristically motivated cost function maximises the energy of the transform coefficients associated with each estimated source. Finally, we evaluate our proposed transform by comparing it against two well-known transforms: the shorttime Fourier transform and the modified discrete cosine transform. We assume that the mixing parameters are known. Our results show that in some cases, our method can give better results than these fixed-basis representations.