Improving Singing Voice Separation With The Wave-U-Net Using Minimum Hyperspherical Energy
Joaquin Perez-Lapillo, Oleksandr Galkin, Tillman Weyde
-
SPS
IEEE Members: $11.00
Non-members: $15.00Length: 15:34
In recent years, deep learning has surpassed traditional approaches to the problem of singing voice separation. The Wave-U-Net is a recent deep network architecture that operates directly on the time domain. The standard Wave-U-Net is trained with data augmentation and early stopping to prevent overfitting. Minimum hyperspherical energy (MHE) regularisation has recently proven to increase generalisation in image classification problems by encouraging a diversified filter configuration. In this work, we apply MHE regularisation to the 1D filters of the Wave-U-Net. We evaluated this approach for separating the vocal part from mixed music audio recordings on the MUSDB18 dataset. We found that adding MHE regularisation to the loss function consistently improves singing voice separation, as measured in the Signal to Distortion Ratio (SDR) on test recordings, leading to the best time-domain system for singing voice extraction.