Harmonic-Temporal Factor Decomposition Incorporating Music Prior Information for Informed Monaural Source Separation

For monaural source separation two main approaches have thus far been adopted. One approach involves applying non-negative matrix factorization (NMF) to an observed magnitude spectrogram, interpreted as a non-negative matrix. The other approach is based on the concept of computational auditory scene analysis (CASA). A CASAbased approach called the “harmonic-temporal clustering (HTC)” aims to cluster the time-frequency components of an observed signal based on a constraint designed according to the local time-frequency structure common in many sound sources (such as harmonicity and the continuity of frequency and amplitude modulations). This paper proposes a new approach for monaural source separation called the “Harmonic-Temporal Factor Decomposition (HTFD)” by introducing a spectrogram model that combines the features of the models employed in the NMF and HTC approaches. We further describe some ideas how to design the prior distributions for the present model to incorporate musically relevant information into the separation scheme.