A frugal approach to music source separation
Published:
During the past years, deep learning brought a big step in performance of music source separation algorithms. A lot has been done on the architecture optimisation, but training data remains an important bias for model comparison. In this work, we choose to work with the frugal and well-known original TasNet neural network and to focus on simple methods to exploit a relatively important dataset. Our results on the MUSDB test set outperform all previous state of the art approaches with extra data on the following source categories: vocals, accompaniment, drums, bass and in average. We believe that our results on how to shape a training set can apply to any type of architecture.
This paper was never published, as it was done at a time where both me and Nathan had little time to give to the project - I was focused on my thesis, and A-Volute (now Steelseries) had hard economic times and were restructuring towards less research tasks. The teaser image in the publications page is from the TasNet paper.