UVR Mel-Band RoFormer vs Demucs โ Music Source Separation Model Comparison
Karaoke-specific 2-stem vs general-purpose 4-stem โ architecture, training, performance, selection criteria
Key Differences at a Glance
| UVR Mel-Band RoFormer Karaoke | Demucs htdemucs | |
|---|---|---|
| Developer | UVR community (aufr33+viperx) | Meta/Facebook AI Research |
| Model size | 913 MB | ~80 MB |
| Output stems | 2 (lead vocals / karaoke accompaniment) | 4 (vocals / drums / bass / other) |
| Vocal definition | Lead vocals only | All vocals (lead + backing) |
| Backing vocals in instrumental | Yes (by design) | No |
| Architecture | Pure frequency-domain Transformer | Time+frequency hybrid Transformer |
| Vocal SDR | 10.20 dB (lead only) | ~9.20 dB (all vocals) |
Architecture Comparison
UVR: waveform โ STFT โ mel-band decomposition (60 bands, 50% overlap) โ alternating time/freq Transformers ร 6 โ mask estimation โ ISTFT
Demucs: waveform โ parallel time-domain (1D CNN) + frequency-domain (STFT + 2D CNN) paths โ Cross-Domain Transformer โ sum both decoders
The Decisive Difference: Training Data
The biggest gap isn't architecture โ it's how training data defines stems.
UVR Karaoke: karaoke = instrumental + backing vocals. other = lead vocals only.
Demucs: vocals = all human voices (lead + backing).
When to Use What
UVR Karaoke: karaoke accompaniment, clean lead vocal extraction, quality-first scenarios.
Demucs: need individual drums/bass/guitar, remixing, instrument practice, model size matters.
Try Both
audio-separator song.mp3 --model_filename mel_band_roformer_karaoke_aufr33_viperx_sdr_10.1956.ckpt
demucs song.mp3
Run on the same song with chorus โ the difference in accompaniment naturalness is immediately obvious.
Key Concepts
UVR outputs 2 stems (lead vocals/karaoke), Demucs outputs 4 stems (vocals/drums/bass/other)
Key difference is training data: UVR "karaoke" stem includes backing vocals
UVR is 913MB (84.2M params), Demucs is ~80MB (26.3M params)
UVR is pure frequency-domain Transformer, Demucs is time+frequency hybrid
Karaoke โ UVR (keeps chorus), remixing โ Demucs (per-part manipulation)