Published November 4, 2025
| Version v3
Model
Open
BSRNN model checkpoints for music separation
Description
Here you can download the pretrained weights (PyTorch Lightning checkpoints) of the BSRNN model for music separation. These are unofficial, and correspond to our implementation, whose code is freely available here. These models are trained using the openly available MUSDB18-HQ dataset.
We provide three sets of pretrained weights:
- bsrnn-large, which corresponds to our attempt at implementatin the original BSRNN model
- bsrnn-opt, which corresponds to our optimized version of BSRNN (see the dedicated code for details)
- simo-bsrnn-opt, which is our optimized version of the SIMO-BSRNN variant.
bsrnn-large and bsrnn-opt consist of 4 checkpoints (one per instrument), while simo-bsrnn-opt is a single checkpoint that can estimates all 4 sources.
The simo-bsrnn-opt model yields the best performance, and it is the lightest overall. However, if you are only interested in extracting a single source (e.g., the vocals), it might be more efficient to use bsrnn-large or bsrnn-opt with the corresponding source checkpoint.
Please check our github repository for instructions on how to use.