The proposed method, an end-to-end lightweight music source separation based on deep learning model using time domain information. The complexity of the model and the computing time are both reduced by introducing Atrous Spatial Pyramid Pooling. The Temporal Attention with Dynamic Convolution Kernel and the Receptive Field Preserving Decoder are proposed to enhance the performance of music source separation.

The proposed single channel end-to-end music source separation model works with image semantic segmentation skills which reducing layer number in network. The proposed model now with only 10% of number of parameters of previously proposed models and shows even better result. The proposed "Depth to Time" also reduced calculation and makes running real-time application on resource limited device possible.

The previously proposed sound source separation models are already claimed with good enough result. However, those models are with high complexity and huge amount of parameters. These models are note suitable for running real-time application on devices with limited resource. Though the proposed method aimed music source and human voice separation, this can be extended to other real-time applications like noise reduction and speech separation.