Monophonic Singing Voice Separation Based on Deep Learning

The traditional monophonic singing voice separation system usually consists of two modules: melody extraction and time-frequency masking. In recent years, with the rapid development of neural networks, end-to-end music separation system that based on deep learning has become more and more popular. Deep neural networks are very useful for processing complex nonlinear data, this paper describes a system based on the framework of the traditional separation system, which uses ResNet to extract the melody of music signals, and combines NMF's soft masking separation algorithm. Compared with the existing module, our separation system is proved that can get better separation effect.