HydraNet: A Network For Singing Voice Separation

Student thesis: Master Thesis and HD Thesis

  • Esbern Torgard Kaspersen
4. term, Medialogy, Master (Master Programme)
This paper proposes a new model, called HydraNet, for solving the problem of single-channel blind source separation. The model is based on two other models called Chimera and Wave-U-Net. By combining these two it was hoped that the signal-to-distortion ratio (SDR) would increase. HydraNet was implemented in Python PyTorch, and evaluated on the DSD100 dataset for singing voice separation. It reached a SDR of 9.78dB for instrument separation and 3.46dB for singing voice separation. Chimera and Wave-U-Net were also implemented in Python PyTorch and tested on DSD100.
Publication date28 May 2019
Number of pages78
ID: 304598915