File size: 1,089 Bytes
7b0ece1
 
 
cb64af1
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
---
license: mit
---
# Bi-SCNet

**Binaural Sparse Compression Network for Music Source Separation**

See: [Spacing Out: On the Reliability of Binaural Music Source Separation Metrics](https://github.com/richa-namballa/binaural-mss/)

**Binaural Model Checkpoint**

* `binaural_scnet.tar.gz`: Checkpoint for the Bi-SCNet model trained on [Binaural-MUSDB](https://github.com/richa-namballa/binaural-mss/tree/main/data) with only RMSE loss.

* `binaural_scnet_spat_loss.tar.gz`: is the checkpoint for the Bi-SCNet model trained on Binaural-MUSDB with the ITD-based spatial loss equally weighted with the RMSE loss.
 The [spatial loss](https://github.com/richa-namballa/Bi-SCNet/blob/d75678e19f969bd4a1c981f794724d4aefb6ef86/scnet/loss.py#L142) is based on ["Interaural Time Difference Loss for Binaural Target Sound Extraction"](https://ieeexplore.ieee.org/abstract/document/10694612) by Hernandez-Olivan et al. (2024)

This model is an adaptation of the [SCNet: Sparse Compression Network for Music Source Separation](https://github.com/starrytong/SCNet) repository by Tong, et al. (2024).