de-leak
A real-time, machine-learning leak gate. It finds content that does not belong in a stem — bleed from other instruments, ghost residue from a separator, RVQ codec aliasing in AI-generated stems — and gates it out, band by band, without flattening the music around it.
01What de-leak does
A conventional gate decides on level. de-leak decides on content: a small neural classifier judges, frame by frame, whether what it is hearing is the stem's own material or something leaking into it.
Split a mix into stems — with a separator, or by asking a generative model for individual parts — and the result is never perfectly clean. The vocal stem carries a shadow of the snare. The "other" stem carries a ghost of everything. AI-generated stems carry something stranger still: non-musical shimmer above roughly 4 kHz that is an artefact of how the model encodes audio, and which no level-based gate will ever catch because it rides along with the music rather than sitting in the gaps.
de-leak works differently:
- A leak classifier — a compact neural network embedded in the plug-in — produces a per-frame gate value from the spectrum: 1.0 means "this is the stem's own material", down toward 0 means "this is leak".
- A three-band multiplier applies that gate with independent strength in the low, mid and high bands, because leak is rarely spread evenly. Vocals bleed differently from cymbals.
- HF compensation decides how much of the high band survives the gate, so you can be aggressive on ghosting without stripping air off the whole track.
- Extra gain reduction deepens the cut in the leak region only, leaving the fully-open regions untouched.
The gate is causal and runs in real time in any DAW, so you can dial it in while listening rather than round-tripping through an offline tool.
de-leak removes material that does not belong. If your problem is instead the smeared, metallic residue inside the material itself — the texture typical of generated audio and low-bitrate codecs — that is de-artifact. AI-generated stems commonly have both problems, and the two are routinely used on the same channel.
02Install
macOS
Download de-leak-0.1.6.pkg from the download page and double-click it. The package is signed and notarized, so there is no Gatekeeper warning. It installs two formats:
| Format | Installed to | Used by |
|---|---|---|
| VST3 | /Library/Audio/Plug-Ins/VST3/de-leak.vst3 | Live, Cubase, Studio One, Reaper, FL Studio… |
| CLAP | /Library/Audio/Plug-Ins/CLAP/de-leak.clap | Bitwig, Reaper, Studio One 6.5+ |
The installer clears the quarantine flag and fixes executable permissions so both formats load cleanly on the first scan.
de-leak ships as VST3 and CLAP only, so it does not appear in Logic Pro or GarageBand, which load Audio Units exclusively. Logic users can run it in a VST3/CLAP-capable host in the meantime. (de-artifact does ship an Audio Unit.)
Windows
Download de-leak-windows-x86_64-0.1.6-setup.exe, or the matching .zip for a manual install. Default locations:
C:\Program Files\Common Files\VST3\de-leak.vst3C:\Program Files\Common Files\CLAP\de-leak.clap
The Windows build is not yet Authenticode-signed — if SmartScreen warns on first run, choose More info → Run anyway.
System requirements
| Item | Requirement |
|---|---|
| macOS | 11 Big Sur or later. Universal — Apple Silicon and Intel. |
| Windows | 10 (1809) or later, 64-bit. |
| Host | Any VST3 or CLAP host. |
| Sample rates | Anything from 8 kHz to 384 kHz — the plug-in resamples to and from the model rate internally. |
| Model | Embedded in the plug-in. No external file, no download step. |
Hosts cache plug-in binaries. After installing a new build over an old one, remove the plug-in from the track and add it again, or restart the DAW — a plain rescan often keeps serving the cached previous version.
03First run
Look for de-leak under vendor Intrect in your host's plug-in list. Insert it on the stem you want to clean.
It reports about 93 ms of latency at 44.1 kHz and the host compensates automatically (§11). On first launch with an internet connection it requests a 14-day trial by itself — see §12.
The plug-in ships with a single, trained gate model built in. There is nothing to choose or load: the behaviour differences you need come from the presets, not from swapping models.
04The interface
A header strip, a wide three-band spectrum analyser you can manipulate directly, and a control sidebar.
Every screenshot on this page is the plug-in processing a real AI-generated track — the curves are its actual output on that material, not an illustration.
4.1 Header
| Element | Behaviour |
|---|---|
| Preset selector | Drop-down listing all seven presets. Choosing one moves every gate control to that preset's values. See §6. |
| Output mode | WET normally, DRY (bypass) when bypassed, DIFF (dry−wet) when you are auditioning the removed signal. Colour-coded so it is unmissable. |
| gate / min | Live gate statistics for the last chunk. gate is the mean gain (1.00 = fully open, nothing being removed) and min is the lowest gain reached. The quickest numeric answer to "is it doing anything, and how hard". |
| LICENSE | Opens the licence panel — status, activation, sign-out. See §12. |
4.2 The spectrum display
| Element | Meaning |
|---|---|
| Grey fill | The input spectrum — what arrives at the plug-in. |
| Cyan line | The output spectrum. The gap between fill and line is the gating; where they overlap, nothing is being removed. |
| Red line | Appears in Diff mode — the removed signal on its own. See §9. |
| Divider lines | The two crossovers between low/mid and mid/high, labelled in Hz. Drag them horizontally. |
| Solo toggles | The circular buttons labelled low, mid, high across the top. |
| Strength handles | One rectangular handle per band, riding a horizontal line at that band's strength value. |
4.3 Control sidebar
Hover any slider for an inline description. All seven gate parameters follow the selected preset until you untick Preset drives params, at which point they become editable and stay where you put them.
05Quick start
- Insert de-leak on the stem you want to clean.
- Pick the preset that matches the source: vocals → Vocal, drums/bass/guitar → Instrument, an AI stem with high-frequency shimmer → AI Ghost, a full mix → Master. Not sure → Default, which is safe on anything.
- Play, and watch the gap between the grey fill and the cyan line. That gap is what is being removed.
- Switch on Diff and listen. You should hear noise, reverb tails and shimmer. If you hear the actual performance, the gate is too aggressive — move to a gentler preset or raise the offending band's strength handle.
- Too gentle instead? Step up to AI Heavy or Aggressive, or untick Preset drives params and tune by hand.
Everything else is refinement. Solo the removed signal, boost your monitoring 20–40 dB if you need to, and ask one question: can I hear the performance in there? If yes, back off. If it is only noise and tails, you have room to push further.
06Presets
Seven presets covering the common stem types. Each one sets all seven gate parameters plus the three band strengths and both crossovers.
| Preset | Use it on | Character | Watch out for |
|---|---|---|---|
| Default Universal |
Any stem, as a safe first pass | Moderate intensity with the midrange protected. Catches obvious leak and little else. | Subtle ghosting survives — by design |
| Vocal | Suno / Udio raw vocal stems | Preserves vocal formants while working the high band harder, where RVQ artefacts live. | Slight loss of HF sparkle on some material |
| Instrument | Drums, bass, guitar, synth | The most conservative preset. Low and mid strengths are pulled well back to protect transients and body. | Leaves subtle leak in place — switch to AI Ghost if you hear shimmer |
| Master | Full mixes, master bus | Nearly transparent. Only egregious leak is touched. | Not a cleanup tool at this setting — it is a safety net |
| AI Ghost | Suno Other / Percussion / some Synth stems | The high-frequency specialist. Crossover moves up to 5.7 kHz, the midrange is strongly protected, and the air band takes a deep cut. Tuned by automated search against percussion and "other" stems plus a safety set. | Cymbal sizzle and hi-hat air are reduced along with the ghosting — judge by ear |
| AI Heavy | Coarse RVQ artefacts on AI vocals | Full intensity with a deep floor and a non-linear gate shape, plus extra gain reduction in the leak region. | Audible damage on clean material — use it where the source is genuinely rough |
| Aggressive | Auditioning, experiments | Boundary protections off, maximum attenuation. Intended for listening to the leak on its own (with Diff), not usually for a final mix. | Not a mixing preset |
Preset values
For reference when tuning by hand — these are the values each preset loads.
| Preset | Intensity | Floor | Crossover | HF comp | Release | Shape | Extra GR |
|---|---|---|---|---|---|---|---|
| Default | 0.60 | −30 dB | 2000 Hz | 0.90 | 0.00 | 1.0 | 0 dB |
| Vocal | 0.80 | −40 dB | 2200 Hz | 0.80 | 0.00 | 1.3 | 3 dB |
| Instrument | 0.40 | −24 dB | 2500 Hz | 0.95 | 0.05 | 1.0 | 0 dB |
| Master | 0.30 | −18 dB | 3000 Hz | 1.00 | 0.10 | 1.0 | 0 dB |
| AI Ghost | 0.60 | −33 dB | 5700 Hz | 0.85 | 0.60 | 1.0 | 0 dB |
| AI Heavy | 1.00 | −60 dB | 2000 Hz | 0.70 | 0.70 | 2.0 | 6 dB |
| Aggressive | 1.00 | −80 dB | 2000 Hz | 0.50 | 0.80 | 3.0 | 18 dB |
And the band settings each preset applies:
| Preset | Low/Mid | Mid/High | Low str. | Mid str. | High str. | High GR |
|---|---|---|---|---|---|---|
| Default | 250 Hz | 4000 Hz | 0.70 | 0.60 | 1.00 | 0 dB |
| Vocal | 200 Hz | 4500 Hz | 1.00 | 0.90 | 1.20 | 3 dB |
| Instrument | 300 Hz | 5000 Hz | 0.50 | 0.50 | 0.90 | 0 dB |
| Master | 200 Hz | 6000 Hz | 0.40 | 0.30 | 0.60 | 0 dB |
| AI Ghost | 250 Hz | 5700 Hz | 0.69 | 0.32 | 1.98 | 29 dB |
| AI Heavy | 250 Hz | 4000 Hz | 1.00 | 0.70 | 1.30 | 6 dB |
| Aggressive | 250 Hz | 4000 Hz | 1.00 | 0.70 | 1.30 | 6 dB |
The AI Ghost row explains its character at a glance: the midrange strength drops to 0.32 while the high band goes to 1.98 with 29 dB of extra gain reduction, all above a crossover pushed up to 5.7 kHz. It is a scalpel aimed at the air band, not a general-purpose gate.
07Band controls
The three bands are always active, and you manipulate them directly on the spectrum rather than through abstract sliders.
7.1 Crossovers
Two vertical divider lines mark the low/mid and mid/high boundaries, each labelled with its frequency. Drag a divider horizontally to move it. The low/mid crossover ranges from 80 Hz to 1 kHz and the mid/high from 1.5 kHz to 12 kHz, both on a skewed scale so the useful region gets most of the travel.
Moving the mid/high crossover is the single most effective adjustment when a preset is nearly right: push it up to protect more of the musical high end, pull it down to bring more of the spectrum under aggressive treatment.
7.2 Strength handles
Each band carries a rectangular handle sitting on a horizontal line, with its value printed underneath.
| Gesture | Effect |
|---|---|
| Drag vertically | Sets band strength, 0 to 2. 0 effectively bypasses the band, 1.0 applies the gate as the model produced it, and 2.0 doubles its effect. |
| Shift + drag | Sets Extra GR for that band, 0 to 40 dB — additional attenuation applied only to the leak portion of that band. Regions the gate considers fully open are not touched. |
| Right-click | Resets: strength back to 1.0, or Extra GR back to 0 if you are in Shift mode. |
They are not the same lever. Strength scales how much the model's decision is trusted in that band — raising it makes borderline material more likely to be gated. Extra GR leaves the decision alone and simply cuts harder wherever the gate has already closed. For "the gate is finding it but not removing enough", reach for Extra GR. For "the gate is not catching it at all", reach for strength.
7.3 Band solo
The circular toggles at the top of each band solo that band. They are multi-select: engage low and high together and only the mid is muted.
Turning them all on, or all off, both mean "no solo" — the full output. This is deliberate, so you can never get stuck in a state where everything is muted.
Solo combines with Diff: with both engaged you hear only the removed signal, and only from the bands you have soloed. That combination is the fastest way to answer "what exactly is this preset taking out of the top end".
08Gate controls
Untick Preset drives params to edit these. While it is ticked they follow the selected preset and are read-only.
| Control | Range | What it does |
|---|---|---|
| Intensity | 0 – 1 | Overall gate strength — how much of the model's decision is applied. The first control to reach for. Typical working range is 0.3 (subtle) to 1.0 (full). |
| Floor (dB) | −90 … −20 | The minimum gain when the gate is fully closed, i.e. how deep the cut is allowed to go. Gentle values around −24 to −30 dB leave a natural bed of residue; −60 dB and below remove leak almost completely but expose any misjudgement by the model. |
| Crossover (Hz) | 200 – 8000 | The pivot for HF compensation below. Sets where "high frequency" begins for the purpose of the next control. |
| HF compensation | 0 – 1 | How much of the high band survives the gate. 1.0 leaves the highs untouched; lower values let the gate work up there too. Lower it when the ghosting is the high band; raise it when cleaning costs you too much air. |
| Release smooth | 0 – 0.99 | How slowly the gate re-closes after opening. 0 is instant, which is precise but can chatter on busy material; higher values ride over short gaps and sound smoother. |
| Gate shape | 1 – 6 | A non-linear exponent on the gate value. 1.0 is linear. Higher values push partially-closed regions much further closed while leaving fully-open regions alone, so the gate becomes decisive rather than gradual. |
| Extra GR | 0 – 40 dB | Global extra attenuation in the leak region, on top of the per-band Extra GR. 0 off, 6 mild, 12 moderate, 24 assertive, 40 brutal. Fully-open regions are never affected. |
Tuning order that works
- Start from the closest preset rather than from scratch.
- Adjust Intensity until Diff contains everything you want gone and nothing you want kept.
- Move the mid/high crossover on the display so the aggressive treatment lands where the problem actually is.
- Use per-band strength handles to protect a band that is taking damage.
- Only then reach for Floor and Extra GR to deepen the cut, and Gate shape if you need the decision to be sharper.
09Diff monitoring
Diff plays dry − wet: exactly what the gate is removing, and nothing else. It is the tool that turns tuning from guesswork into a decision.
DIFF (dry−wet) and the removed signal is drawn in red — concentrated in the high band, which is the shape you want to see.
DRY (bypass) and the output trace sits on top of the input fill — no gap between them means nothing is being removed. This is your A/B against the two shots above.How to listen
- Boost your monitoring by 20–40 dB. The removed signal is quiet by design, and at normal level you will conclude it is empty when it is not.
- If you hear the performance — words, melody, drum hits with their body intact — the gate is over-reaching. Reduce intensity, or raise the strength handle of the band it is coming from.
- If you hear noise, reverb tails, shimmer and hash — it is doing its job, and you may have room to push further.
- Combine Diff with band solo to isolate one region at a time.
A gate that removes leak will also take reverb tails with it, because to a classifier they look similar — energy that continues after the source event. Whether that is damage or cleanup depends on what you are making. On a dry pop vocal it is usually a gain; on an ambient piece it may be exactly what you do not want. Diff lets you hear the trade rather than discover it later.
10Workflow recipes
10.1 Making a set of AI stems usable
- Insert de-leak on every stem.
- Match preset to stem type — lead vocals → Vocal; drums and percussion → AI Ghost; bass → Instrument (protects the low end); guitar and synth → Instrument or AI Ghost; the "other" stem → AI Ghost, which is usually the worst offender.
- Go through them one at a time with Diff on and confirm each is removing residue rather than material.
- Where a stem takes damage, raise that band's strength handle instead of abandoning the preset.
- Optionally add one instance with the Master preset on the master bus as a final safety pass.
10.2 Chasing high-frequency shimmer
- Solo the high band and play the track.
- If you hear music — cymbals, air, sibilance — the top end is fine and the problem is elsewhere.
- If you hear hiss or a metallic shimmer that does not track the music, that is RVQ ghosting: load AI Ghost.
- Turn Diff on to confirm what would be removed before committing.
- If it takes too much cymbal with it, raise the mid/high crossover or reduce the high band's Extra GR.
10.3 Taming a reverb tail on vocals
- Load Vocal or AI Heavy depending on how much tail you want gone.
- Raise Release smooth so the gate does not chatter between syllables.
- With Diff on, confirm you are hearing tail rather than the vocal core.
- Solo the mid band to check the body of the voice specifically — that is where damage shows up first.
10.4 Using it with de-artifact
When a stem has both problems, order matters less than monitoring, but a good default is de-leak first, de-artifact second: remove the material that does not belong, then clean the texture of what remains. Check each stage on its own — solo one plug-in's monitor at a time rather than trying to judge the pair at once.
11Latency & performance
| Item | Value |
|---|---|
| Reported latency | 4096 samples ≈ 93 ms @ 44.1 kHz |
| Other sample rates | Scales with the host rate, so the figure in milliseconds stays about the same. |
| Compensation | Reported to the host, which lines tracks back up automatically. |
Every major DAW applies plug-in delay compensation without configuration. If an offline bounce comes back silent or truncated, PDC is the first thing to check in the host's settings.
CPU
The gate model is small — inference is a negligible share of the total cost. Most of the work is the FFT analysis and resynthesis, which is fixed per chunk regardless of settings. Practical consequences:
- Changing presets or pushing controls harder does not meaningfully change CPU usage.
- Very small host buffers can cause spikes, because the fixed per-chunk work has less room to spread out. Raising the buffer size is the fix.
- Running de-leak on many stems at once is normal usage and scales linearly.
12Licence
A 14-day free trial, then a paid licence. Nothing is gated at install time.
On first launch with internet access the plug-in requests a 14-day trial automatically, bound to a one-way fingerprint of your machine and stored as a signed licence file on disk. After that no connection is needed for the rest of the trial. If the server cannot be reached, a local fallback trial is issued so audio is never blocked, and the plug-in retries in the background.
| State | Meaning | Audio |
|---|---|---|
| Activated | Paid licence bound to this machine. | Full output, no expiry. |
| Trial · N days | Free evaluation, fully functional. | Full output. |
| Trial expired | The evaluation period is over. | A 50 ms silent gap every 10 seconds until you activate. Everything else keeps working. |
| Needs refresh | A paid licence is past its offline refresh window. | Gapped until the next successful background refresh — open the editor while online. |
Activating
- Buy a licence; the key arrives by email.
- Open the licence panel from the LICENSE button in the header.
- Paste the key and click Activate. The server signs a licence bound to this machine and the plug-in installs it.
- It refreshes itself in the background from then on. If the machine has been offline for a long stretch, the next refresh catches it up.
One licence covers three machines. If you hit the limit, the panel lets you release one of the existing seats. Moving to a new computer: Sign out on the old one first, then activate on the new one.
Files and privacy
- macOS —
~/Library/Application Support/Intrect/de-leak-rt/license.blob - Windows —
%APPDATA%\Intrect\de-leak-rt\license.blob
The server stores a non-reversible machine fingerprint, the plug-in name, and trial/activation timestamps. No audio, no telemetry, no usage statistics. The plug-in contacts the server only for licence events: first-launch trial, activation, refresh and sign-out.
13Troubleshooting & FAQ
- de-leak does not appear in Logic Pro.
- Logic loads Audio Units only, and de-leak currently ships as VST3 and CLAP. Use a VST3- or CLAP-capable host for now. de-artifact does ship an Audio Unit if you need something in Logic today.
- An offline bounce is silent or truncated.
- The host must apply plug-in delay compensation for the roughly 93 ms the plug-in reports. Every major DAW does this by default; if yours has a PDC setting, make sure it is on. If it still happens, open a ticket with your host and version.
- Does it work at sample rates other than 44.1 kHz?
- Yes. The plug-in resamples between the host rate and the model rate internally, anywhere from 8 kHz to 384 kHz. Nothing to configure.
- I hear clicks when toggling solo.
- Solo gain is ramped rather than switched, so this should not happen on a current build. If you still hear clicks, please report it with your host and buffer size.
- With Diff and solo both on, the output sounds like the dry input.
- That was a bug in older builds where solo was applied to only one path. On current builds solo applies to both the wet and dry paths, so muted bands cancel and only the active band's difference survives. Update if you see this.
- CPU usage seems high on small buffers.
- The dominant cost is the fixed FFT analysis and resynthesis per chunk, not the neural network. Raise the host buffer size — that is almost always the fix.
- The gate removes reverb I wanted to keep.
- Expected behaviour with an aggressive preset: tails and leak look alike to the classifier. Move to a gentler preset, raise the affected band's strength handle, or increase Release smooth so the gate holds open through the tail.
- It is not catching the ghosting I can clearly hear.
- Solo the high band to confirm where it lives, then load AI Ghost, which is built for exactly that material. If it is still surviving, untick Preset drives params and lower HF compensation so the gate is allowed to work in the high band at all.
- What is the difference between de-leak and de-artifact?
- de-leak removes content that does not belong in the stem — bleed and ghosting. de-artifact removes the smeared, metallic texture within the material, typical of generated audio and lossy codecs. Different problems, frequently present together.
14Support
Open a ticket at intrect.io/support. Include your host, operating system, sample rate and buffer size, plus which preset you were using — that combination usually narrows a report down immediately.
- Platforms — macOS 11+ (Apple Silicon and Intel) and Windows 10+ (x86_64), as VST3 and CLAP. Linux is not packaged.
- Signing — the macOS package is signed and notarized; the Windows installer is not yet Authenticode-signed.
Also worth reading
de-artifact user manual · de-leak product page · Changelog · Help centre