Structure-aware dictionary learning with harmonic atoms

Non-negative blind signal decomposition methods are widely used for musical signal processing tasks, such as automatic transcription and source separation. A spectrogram can be decomposed into a dictionary of full spectrum basis atoms and their corresponding time activation vectors using methods such as Non-negative Matrix Factorisation (NMF) and Non-negative K-SVD (NN-K-SVD). These methods are constrained by their learning order and problems posed by overlapping sources in the time and frequency domains of the source spectrogram. We consider that it may be possible to improve on current results by providing prior knowledge on the number of sources in a given spectrogram and on the individual structure of the basis atoms, an approach we refer to as structure-aware dictionary learning. In this work we consider dictionary recoverability of harmonic atoms, as harmonicity is a common structure in music signals. We present results showing improvements in recoverability using structure-aware decomposition methods, based on NN-K-SVD and NMF. Finally we propose an alternative structure-aware dictionary learning algorithm incorporating the advantages of NMF and NN-K-SVD.