Unsupervised Approach to Music Source Separation using Generalized Dirichlet Prior

Music source separation aims to extract and reconstruct individual instrument sounds that constitute a mixture sound. It has received a great deal of attention recently due to its importance in the audio signal processing. In addition to its stand-alone applications such as noise reduction and instrument-wise equalization, the source separation can directly affect the performance of the various music information retrieval algorithms when used as a pre-processing. However, conventional source separation algorithms have failed to show satisfactory performance especially without the aid of spatial or musical information about the target source. To deal with this problem, we have focused on the spectral and temporal characteristics of sounds that can be observed in the spectrogram. Spectrogram decomposition is a commonly used technique to exploit such characteristics; however, only a few simple characteristics such as sparsity were utilizable so far because most of the characteristics were difficult to be expressed in the form of algorithms. The main goal of this thesis is to investigate the possibility of using generalized Dirichlet prior to constrain spectral/temporal bases of the spectrogram decomposition algorithms. As the generalized Dirichlet prior is not only simple but also flexible in its usage, it enables us to utilize more characteristics in the spectrogram decomposition frameworks. From harmonicpercussive sound separation to harmonic instrument sound separation, we apply the generalized Dirichlet prior to various tasks and verify its flexible usage as well as fine performance.