Music Source Separation Using Deform-Conv Dense U-Net

Music source separation is a research field to separate a target sound source from a sound source consisting of multiple instruments. Convolutional Neural Networks (CNNs) consisting of convolutional layers are the mainstream methods for music source separation. However, these conventional methods have a problem of degradation in separation performance at points where the time-frequency characteristics change rapidly. The problem has to do with the different shapes of the in-strument's sound on the spectrogram. In this paper, we focused on the relationship between musical instrument sounds andex-isting CNN to improve performance. As mentioned above, there are differences in the spectrograms of each instrumental sound, but the conventional methods ignore the differences and process them uniformly. For this reason, we considered that the appropriate convolution might not be applied to each instrument sound. Therefore, we thought that we could improve the separation performance by deforming the shape of thegazing region according to the local feature changes in the time-frequency space of the music component sounds. Specifically, we propose a method using Deformable Convolution (Deform-Conv) to follow the local features of each instrumental sound. Compared with the conventional method, the SDR (separation performance measure) of Vocals was improved by 0.4 pt, confirming the basic effectiveness of this method.