A re-implementation of the (unofficial) SCNet audio source separation model using Pytorch Lightning to facilitate debugging and reproduction.
You can find the pure pytorch implementation in https://github.com/amanteur/SCNet-PyTorch and the original model description in this paper.
While I was unable to reach the same metrics as the original paper, they're still pretty good. The experiments were with the MUSDB18-HQ dataset, using the recommended partitions.
| SDR | |
|---|---|
| Vocals | 9.44 |
| Bass | 8.14 |
| Drums | 10.05 |
| Other | 6.40 |
| Overall | 8.48 |
- Train with additional data.
- Write a proper tutorial.
- Create an end-user script for audio separation.