Note Our reproducible MUSDB18-HQ benchmark with SDR/ISR/SIR/SAR for 5 SOTA models. v1.1 (May 2026).

Politrees/UVR_resources

Updated Dec 23, 2025 • 49

Note UVR community model resources, including the MDX-Net checkpoints we benchmark (e.g. UVR_MDXNET_KARA_2 as `mdx_net_inst_hq3`).

anvuew/BS-RoFormer

Updated Apr 17 • 173 • 6

Note Band-Split Roformer — current SOTA for music source separation. We will add it to v1.2 of our benchmark once we run on CUDA.

anvuew/dereverb_bs_roformer

Updated Feb 13 • 235 • 21

Note BS-Roformer fine-tuned for dereverb. Useful for cleaning up vocal stems before downstream tasks.

mlx-community/demucs-mlx

Audio-to-Audio • Updated Mar 16 • 6

Note HT-Demucs converted to Apple's MLX framework. Native Apple Silicon inference without PyTorch/MPS overhead.

HiDolen/Mini-BS-RoFormer-V2-46.8M

Audio-to-Audio • 46.8M • Updated Dec 30, 2025 • 74 • 5

Note Compact 46.8M-parameter BS-Roformer variant. Lower latency than the full model — useful when you can trade ~1 dB SDR for ~3× speedup.

csukuangfj/spleeter-torch

Updated May 22, 2025 • 8

Note Deezer's Spleeter ported to PyTorch. Historical baseline; useful for reproducing older papers without setting up TensorFlow.

Eddycrack864/Music-Source-Separation-Training

Updated Jun 17, 2024 • 17

Note Reference training scripts and configs used by the source-separation research community.

StemSplitio/htdemucs-ft-pytorch

Audio-to-Audio • Updated 5 days ago

Note Our PyTorch packaging of HT-Demucs FT, ready for HF Inference Endpoints. Highest vocal SDR (9.19 dB) in our benchmark.