Search academic texts

Information × Registration Number 2124U004700, Article popup.category Препринт Title popup.author Svystun Taras popup.publication 01-01-2024 popup.source_user Український католицький університет popup.source https://hdl.handle.net/20.500.14570/4849 popup.publisher Description The field of music source separation aims to split a musical composition into 4 groups of instruments: vocals, drums, bass and others. Increasingly often, researchers use remixing and data augmentation to enlarge the amount of data and its diversity. The purpose of this thesis is to investigate what performance gains can be expected from remixing, as well as what optimal parameters are required for this. Therefore, this study uses the classical remixing method and tests a new method, which consists of mixing human speech with musical accompaniment. After finding the optimal hyperparameters, another data augmentations are used during the final training. The results of this methodology are the list of optimal parameters for remixing, all the steps necessary to reproduce the results, and the improved checkpoint for the SOTA MSS model - Band-split RNN, previously best publicly available. The findings indicate that remixing and data augmentation techniques are indeed powerful and necessary to improve vocal source separation. popup.nrat_date 2025-05-09 Close