Institutional Repository of Vilnius University
Vokalų ištraukimas iš muzikinių failų /
Abstract
dc:descriptionThis thesis examines the problem of vocal extraction from music recordings using artificial intelligencebased models. The aim of the work is to design, train, and compare several neural ne twork architectures for separating the vocal signal from the accompaniment. The thesis discusses several neural network architectures: Unet, BS-RoFormer, MB-RoFormer, ”Windowed-sink” MB-RoFormer, and Mamba blocks. Experiments were conducted with four architecture groups with variations: Unet with trans former blocks (three variations), ”Windowed-sink” MB-RoFormer, MB-RoFormer (two variations), and MB-RoFormer with Mamba blocks (two variations). The models were trained using the MUSDB18 dataset and evaluated according to the SDR, SI-SDR, SIR, SAR, and ISR metrics.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Belevičius, Benediktas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308337124&prefLang=en_US