You ran a mix through Demucs, htdemucs or Spleeter, pulled out the vocal, and it is not clean. Something is wrong with it, but it is hard to name: a bit glassy, a bit smeared, maybe a ghost of the snare where there should be silence between phrases.
That is usually two problems wearing one coat, and they do not respond to the same treatment. Getting them apart is most of the work.
A separated stem carries two distinct kinds of damage. Residue is
harm done to the signal you kept — a smeared, slightly metallic layer sitting
mostly above 3 kHz, plus ringing around transients and sustained tones.
Bleed is signal that belongs to a different stem, audible in this
one. Residue is subtracted by a model that estimates, per frequency bin, how much of
the energy there is artifact rather than music — that is
de-artifact, on the Vocal preset.
Bleed is gated by a classifier that decides, frame by frame and band by band, whether
what it is hearing belongs here — that is
de-leak. Most separated stems have some of both,
and the two plug-ins are frequently used on the same source.
| Residue | Bleed | |
|---|---|---|
| What it is | Damage to the part you wanted to keep. | Content that belongs to another stem. |
| What it sounds like | Glassy or watery sustains, a metallic sheen on consonants, ringing that follows transients. | A recognisable ghost — a snare hit, a bass note, a chord — where the vocal should be alone. |
| Where it sits | Spread across the spectrum, concentrated in the top octaves. | Wherever the source instrument sat. Often low and mid. |
| The tool | de-artifact — subtracts an estimate of the artifact per frequency bin. | de-leak — a leak gate with an ML classifier (F1 = 0.993) that reduces gain only where leakage is detected. |
Solo the stem and listen to the gaps — the moments between phrases where the singer is not singing. That is where the two separate cleanly:
Order matters. Gate the bleed first, then clean what is left. Running de-artifact on a stem that still has drums in it asks the model to decide whether a snare is an artifact — it is not, and that is not a question it was trained to answer.
Vocal preset
It is harmonic-weighted and transient- and sibilance-aware, which is what a narrowband source needs: the same settings that work on a full mix will treat consonants as artifacts. Loading a preset sets every control at once, and every control stays free afterwards.
Strength runs 0–200 %, defaulting to 60 %. 100 % is the model's nominal estimate; above that it deliberately over-subtracts. There is no correct number — there is a point past which you start removing the singer, and the next step is how you find it.
Residual monitor
Residual plays only what is being subtracted. You want shimmer and hash. If you can hear vocal body, breath or the shape of consonants in it, you are over-processing. This is the single check worth doing on every stem.
de-artifact already measures the high-frequency energy it removed and restores it automatically; Brightness adds to or subtracts from that restoration. Reaching for an EQ after the fact is treating a symptom the plug-in already has a control for.
Separation is the median-filter kernel that splits harmonic from percussive, and it has two positions: Normal (17) and Strong (31). Strong separates the two channels more cleanly, so the Harmonic and Percussive strengths land more precisely — at slightly higher CPU cost.
de-artifact is not built for this and will not fix it. A leak gate decides on content rather than level: a small classifier judges, frame by frame, whether what it is hearing belongs to this stem, and applies multiband gain reduction only where it does not. That is de-leak, and it works on vocal stems, instrument stems and any Demucs, htdemucs or Spleeter output.
Logic and GarageBand users: de-leak ships as VST3 and CLAP only, so it does not appear in hosts that load Audio Units exclusively. de-artifact does ship an Audio Unit. In the meantime de-leak runs in any VST3- or CLAP-capable host.
The reduction curve is measured from the actual output, so it moves with the music. On a vocal stem you should see it dip in the top octaves and stay near flat lower down. That shape — deep up high, quiet below — is what healthy artifact removal looks like. Deep reduction low down is unusual: check the Residual, and if you can hear vocal fundamentals in it, lower Strength.
A near-flat curve is a result, not a fault. If a stem came out of the separator in good shape, there is not much to remove, and reading a flat curve as a broken plug-in and cranking Strength is the most common way to damage a track.
Clean a 30-second preview in the browser — no account, nothing to install.
Residue is damage to the signal you kept: a smeared, slightly metallic layer the separator leaves behind, concentrated in the upper spectrum, plus ringing around transients and sustained tones. Bleed is content that belongs to a different stem entirely — drums audible inside the vocal. de-artifact subtracts residue; de-leak gates bleed. They are frequently used on the same source.
Vocal. It is harmonic-weighted and transient- and sibilance-aware, and it is the preset the manual points to for separated vocal stems. Load it, then adjust Strength from its 60% default rather than reaching for anything else first.
Engage the Residual monitor, which plays only what is being subtracted. You want shimmer and hash. If you can hear vocal body, breath or consonants in it, you are over-processing — lower Strength until you cannot.
Not yet. de-leak ships VST3 and CLAP only, so it does not appear in Logic Pro or GarageBand, which load Audio Units exclusively. de-artifact does ship an Audio Unit. Logic users can run de-leak in a VST3- or CLAP-capable host in the meantime.
You can clean a 30-second preview in the browser at try.intrect.io with no account. The plug-in itself has a 14-day trial with every preset and parameter unlocked.