Audio Source Separation by Time-Frequency Masking using a Signal-Adaptive Local Cosine Transform

Audio source separation of instantaneous, two-channel mixtures by time-frequency masking depends on (approximately) disjoint representations of the sources in some transform domain. We investigate the application of cosine packet (CP) trees to perform this transform. A computationally efficient best basis algorithm is applied to trees of local cosine bases to determine an appropriate transform. We concentrate on demixing the sources by binary masking, and assume the mixing parameters are known. We develop a heuristically motivated cost function which maximises the energy of the transform coefficients associated with a particular source. Finally, we evaluate our proposed transform method by comparing it against more well-known transforms such as the short-time Fourier transform and modified discrete cosine transform. It is shown that in some circumstances, our method of adaptively selecting local cosine bases can give better results than fixed-basis representations. ∗ Corresponding author. Email addresses: andrew.nesbit@elec.qmul.ac.uk (Andrew Nesbit), mark.plumbley@elec.qmul.ac.uk (Mark D. Plumbley), mike.davies@ed.ac.uk (Mike E. Davies). 1 Supported by the Department of Electronic Engineering, Queen Mary, University of London, and the Semantic Interaction with Music Audio