Singing Voice Separation Based on Deep Regression Neural Network
Shuqian Yang, Wei-Qiang Zhang · 2019
The goal of singing voice separation is to separate the vocals and the accompaniments from a monaural music mixture. In this paper, we apply the speech enhancement framework to the singing voice separation task and make some improvements on the architecture of the model. We utilize a nonlinear regression function based on deep neural network architecture to extract accompaniments. We extract the log-power spectral features of music and use five ReLU layers in the regression mode. The experimental results in DSD100 dataset show that our method can achieve good results in the field of singing voice separation.