Concept · Model benchmark
BS-RoFormer vs Demucs: Which Source Separation Model Should You Use?
Three architectures dominate source separation. We run one of them in production — so we can give you the comparison straight, including where our own model family loses.
Three architectures, briefly
| Model | Architecture | Best at |
|---|---|---|
| Spleeter | Spectrogram U-Net | Fast, lightweight, older quality |
| Demucs (HTDemucs) | Hybrid waveform-spectrogram U-Net + transformer | Strong all-round open source |
| BS-RoFormer | Band-split RoPE Transformer | Highest SDR, especially bass |
Quality (SDR) — community-aligned numbers
These align with public community benchmarks and model cards; we cite them because we stand behind the same figures.
| Model | Typical SDR | Note |
|---|---|---|
| Spleeter | ~5.4 dB | Baseline |
| htdemucs | ~7.7–8.5 dB | Open-source SOTA for years |
| BS-RoFormer | ~9.8 dB | Largest gap on bass (~11.28 dB) |
Speed, VRAM, cost
| Spleeter | Demucs | BS-RoFormer | |
|---|---|---|---|
| Speed (CPU) | Fast | Slow | Slowest |
| VRAM (4-stem) | ~1–2 GB | ~4–6 GB | ~6–10 GB |
| Setup | pip | CUDA + weights | CUDA + weights + tuning |
Mel-band vs BS-RoFormer
The RoFormer family also has mel-band variants that trade a little quality for lower compute. For delivery-grade separation, the band-split full model is the one that hits the ~9.8 dB range.
What teams actually do
The community pattern: Demucs or MDX for fast batch drafts, then the higher-SDR model for final delivery. Our API collapses that into one pipeline — draft and final both available as presets.