The paper presents Mel-RoFormer, a model for music source separation. It adopts the Mel-band scheme that maps frequency bins into overlapped subbands according to the mel scale. This model has shown superior performance over BS-RoFormer in music separation tasks using the MUSDB18HQ dataset. The previous model, BS-RoFormer, used a band-split scheme defined empirically, but Mel-RoFormer’s scheme is based on the mel scale, a fundamental reference for acoustic feature design in audio signal processing. This new model achieves better results in separating vocals, drums, and other stems in music.
Publication date: 4 Oct 2023
Project Page: https://bytedance.com/mel-roformer
Paper: https://arxiv.org/pdf/2310.01809
