The study of separating the vocal from the accompaniment in single-channel music is foundational and critical in the field of music information retrieval (MIR). Mainstream music-separation methods are usually based on the frequency-domain characteristics of music signals, and the phase information of the music is lost during time–frequency decomposition. In recent years, deep learning models based on speech time-domain signals, such as Conv-TasNet, have shown great potential. However, for the vocal and accompaniment separation problem, there is no suitable time-domain music-separation model. Since the vocal and the accompaniment in music have a higher synergy and similarity than the voices of two speakers in speech, separating the vocal and accompaniment using a speech-separation model is not ideal. Based on this, we propose VAT-SNet; this optimizes the network structure of Conv-TasNet, which sets sample-level convolution in the encoder and decoder to preserve deep acoustic features, and takes vocal embedding and accompaniment embedding generated by the auxiliary network as references to improve the purity of the separation of the vocal and accompaniment. The results from public music datasets show that the quality of the vocal and accompaniment separated by VAT-SNet is improved in GSNR, GSIR, and GSAR compared with Conv-TasNet and mainstream separation methods, such as U-Net, SH-4stack, etc.

Time-domain signal models have been widely applied to single-channel music source separation tasks due to their ability to overcome the limitations of fixed spectral representations and phase information loss. However, the high acoustic similarity and synchronous temporal evolution between vocals and accompaniment make accurate separation challenging for existing time-domain models. These challenges are mainly reflected in two aspects: (1) the lack of a dynamic mechanism to evaluate the contribution of each source during feature fusion, and (2) difficulty in capturing fine-grained temporal details, often resulting in local artifacts in the output. To address these issues, we propose an attention-driven time-domain convolutional network for vocal and accompaniment source separation. Specifically, we design an embedding attention module to perform adaptive source weighting, enabling the network to emphasize components more relevant to the target mask during training. In addition, an efficient convolutional block attention module is developed to enhance local feature extraction. This module integrates an efficient channel attention mechanism based on one-dimensional convolution while preserving spatial attention, thereby improving the ability to learn discriminative features from the target audio. Comprehensive evaluations on public music datasets demonstrate the effectiveness of the proposed model and its significant improvements over existing approaches.