live-coding-academy_ ⌕ search
home/ atoms/ sdr-source-to-distortion-ratio

SDR (signal-to-distortion ratio) is the standard metric for evaluating stem separation quality

Concept L3Craft K · AI & real-time generative AV

Signal-to-Distortion Ratio (SDR), measured in decibels, quantifies how cleanly a model isolates one stem. Higher SDR means less bleed from other instruments and fewer artefacts. Benchmarks use the MUSDB or MUSDB-HQ dataset — a set of professionally mixed multi-track recordings released for research. A simple waveform-only model (Wave-U-Net) achieves ~3.2 dB; Hybrid Transformer Demucs v4 achieves 9.0 dB on MUSDB-HQ. Perceptual quality ratings (MOS) for naturalness and contamination are reported alongside SDR, because a high-SDR stem can still sound unpleasant. When choosing a Demucs model variant for a practical task, SDR is the primary selection criterion.

Examples

Model comparison from the README:

  • Wave-U-Net (waveform only): 3.2 dB SDR
  • Hybrid Demucs v3 (no extra data): 7.7 dB SDR
  • HT Demucs v4 fine-tuned (800 extra songs): 9.0 dB SDR

Assessment

What does a 2 dB improvement in SDR mean perceptually? At what SDR level are stems typically usable as sampling material without significant clean-up?

Prerequisites — understand these first

Don't confuse with

Sources

“`Overall SDR` is the mean of the SDR for each of the 4 sources, `MOS Quality` is a rating from 1 to 5 of the naturalness and absence of artifacts given by human listeners”
corpus · demucs-music-source-stem-separation · chunk 5

Taught in modules

Unlocks (1) — atoms that build on this