Stem Separation Quality: Models, SDR, and High-Mid Artifacts
Spectrogram models and hybrid waveform models fail in different places. A null test shows whether high-mid transients survived. SDR from a paper is not a score for the file in front of you.
Overview
MixMaster Pro is the living Mix Intelligence platform for producers. Analyses stay in your Mix Library, revisions compare on the same mix, references save to your account, and Maya on Pro+ coaches from your bound report - connected workflows instead of one-time uploads.
A published SDR number answers a lab question: how close was the estimate to a hidden ground-truth stem on a dataset such as MUSDB18. Your bounce has no hidden stem. The useful check is a null test against the original mix, especially in the high-mids where consonants, cymbals, and pick noise live.
MixMaster Pro runs one stem job and returns downloadable parts. It does not expose a model picker, and it does not print an SDR score. Judge the files, then keep the original bounce wherever the null is loud.
Related topics
What an SDR benchmark actually measures
Signal-to-distortion ratio is in decibels. Higher means the estimate is closer to a known target stem, after the usual scale alignment. Papers report a mean across songs. That mean hides the one chorus where the vocal transient turned to noise.
- SDR needs ground truth. A commercial master does not have it, so you cannot compute a real SDR on that file.
- A 1 dB paper gap between model families is smaller than the gap between a clean WAV and a crushed MP3 of the same song.
- SDR does not score taste. A stem can score well and still sound watery on a hi-hat.
- Do not treat a blog's "we scored 9 dB" table as a measurement unless they name the dataset, the metric, and the exact checkpoint.
Two architectures, two failure modes
MDX-Net-style models estimate a mask on a spectrogram, in the family popularized by the Music Demixing Challenge and tools such as Ultimate Vocal Remover. Hybrid Demucs-style models (including HTDemucs) mix a waveform network with a spectrogram network. Meta published that hybrid line. Neither class wins every song.
- Spectrogram masks often isolate a pop vocal cleanly, then smear the attack. The leftover is musical noise between about 2 and 8 kHz.
- Hybrid waveform models often keep drum and bass attacks more intact, then leave cymbal or reverb in the vocal, or a soft chorus on tails.
- Ensembles (run both, keep the stem whose null is quieter in the band you care about) are a listening decision. MixMaster Pro does not run that bake-off for you.
- Bleed and metallic tails are covered as listening checks in the stem artifacts article. This page is the measurement behind those checks.
Null test for high-mid transient loss
Sum the separated stems at unity gain, with no extra fader moves. Polarity-invert the original bounce and play it against that sum. What you hear is the error: missing energy, doubled energy, and model residue.
- 1. Keep the original bounce on a track. Do not master it or change its gain.
- 2. Import every stem. Set each fader to unity and bypass processing.
- 3. Invert the polarity of the original only. Solo the original plus the stem sum.
- 4. If the sum is quieter than either file alone, the split mostly reconstructed the mix. A loud null means level or polarity is wrong - fix that before you judge tone.
- 5. Band-pass or solo roughly 2-8 kHz on the null. Loud consonants, hat attacks, or pick noise there were not preserved.
- 6. A quiet high-mid null means those transients survived. A watery, flanging null means the model smeared them.
- 7. Write down which stem carries the error. Process that stem less, or leave that band on the original bounce.
What to do with a bad split
- Use the stem for the edit that stayed clean (a vocal ride, a drum replace) and leave the damaged band on the original.
- Do not "fix" a metallic high-mid by adding air EQ to every other stem. You stack the same error.
- Re-export from a WAV or AIFF if the source was a low-bitrate MP3. Compression artifacts become separation artifacts.
- Upload the full mix - not the stem sum - when you want MixMaster Pro to score balance. Stem Insights (Studio and above, or Credits) uses stems as probes. The Stems tool is the download.
Key takeaways
- SDR compares a model to hidden ground truth. Your master has none, so use a null test.
- Listen to the null between 2 and 8 kHz for lost consonants, cymbals, and pick noise.
- MixMaster Pro returns one stem job. It does not publish a model leaderboard.
Try MixMaster Pro free
Upload a stereo mix and get AI-powered analysis with prioritized action items in about 60 seconds. No plugin install required.
Start Free Analysis