RTrans-D3Net: Music Source Separation with Multi-Dilated Convolutions and Transformer

Music source separation aims to isolate individual sound sources from mixed music signals. To address this task, we propose a novel deep learning model, RTransD3Net. The model is based on an encoder-decoder architecture and innovatively integrates D3 Block, Res Path, and Transformer modules to overcome the limitations of existing methods, such as restricted receptive fields, feature semantic gaps, and insufficient long-term dependency modeling. Specifically, the D3 Block expands the receptive field and avoids aliasing effects through densely connected multi-dilated convolutions; Res Path reduces feature discrepancies between the encoder and decoder by introducing convolutional layers; and the Transformer module embedded in the bottleneck enhances long-term contextual modeling capabilities. Experiments on the MUSDB18 dataset demonstrate that RTrans-D3Net achieves an average SDR of 6.09 dB, outperforming some existing time-domain and frequency-domain separation methods in terms of separation performance.