Sams-Net: A Sliced Attention-based Neural Network for Music Source Separation

This webpage is to show some listening examples for our Sams-Net method and the state-of-the-art methods Open-Unmix[1] and Demucs[2]. It enables different contributions of feature interactions, achieves easier parallel computing and has a larger receptive field compared with LSTMs and CNNs respectively. Experiments show that the proposed framework outperforms most of the state-of-the-art DNN-based methods with fewer parameters.

Music chunks collected from the Internet

Track:

  1. A-R: Adele - Rolling in the Deep
  2. C-Y: Carpenter - Yesterday Once More
  3. S-J: Sarah Connor - Just One Last Dance
  4. T-S: The Walkers - Sha-La-La-La-La
  5. J-T: Jascha Richter - Take Me To Your Heart
  6. T-E: The Chainsmokers - Everybody Hates Me
Track A-R C-Y S-J T-S J-T T-E
Mixture Input
SamsNet:Vocals
SamsNet:Drums
SamsNet:Bass
SamsNet:Other
OpenUnmix:Vocals
OpenUnmix:Drums
OpenUnmix:Bass
OpenUnmix:Other
Demucs:Vocals
Demucs:Drums
Demucs:Bass
Demucs:Other

Reference

[1] F.-R. Stoter, S. Uhlich, A. Liutkus, and Y. Mitsufuji, "Open-unmix-a reference implementation for music source separation," Journal of Open Source Software, 2019.

[2] A. Defossez, N. Usunier, L. Bottou, and F. Bach, "Music Source Separation in the Waveform Domain," HAL, Tech. Rep. 02379796v1, 2019.