de-leak-rt
Take the bleed out of a split vocal. For vocal stems from Demucs, Suno and Udio splits. Two engines, Gate and Leak Clean, take out the drums and instruments left under the voice, and Residual plays exactly what came out.
For mixing and bouncing, not live monitoring. Not a tool for passing AI detectors.
Hear it on a split vocal
0:00 / 0:00Ay, Corazónvocal stem
Removed: plays at the original's gain. Listen for any part of the voice that went with the bleed.
Gate, Leak Clean, or both. In your DAW.Hear it on a real stemRead the manual
Specifications
Specifications last reviewed against the shipping build, 0.2.2.
Full documentation — every control, engine and preset, with screenshots of the 0.2.1 window.
Product details and community feedback on KVR Audio (opens in a new tab)
- Formats
- VST3 · CLAP · AU · AAX (Apple Silicon) macOS: VST3, CLAP and AU (Apple Silicon and Intel), plus AAX for native Pro Tools on Apple Silicon (Gate and Leak Clean; Intel and Rosetta Pro Tools are not supported). Windows: VST3 and CLAP. No standalone application.
- Engines
- Gate · Leak Clean · Gate → Clean A learned 16-band leak gate, a 64-band direct-sound model that runs on the vocal alone, or the two in series. How they differ
- Latency
- ≈ 257 ms at 44.1 kHz In the Standard setting: 11,334 samples, fixed for the instance and reported to the host, so delay compensation lines the session back up. The Low latency setting leaves out the HF Restore stage and reports less. For mixing and bouncing, not live monitoring.
- Presets
- 8 factory Medium is the default. A preset writes its values into the controls, which stay editable; save your own with Save As. See all 8
- Platform
- macOS 11+ (Universal) · Windows 10 x64 Apple Silicon + Intel in one .pkg. Stereo plug-in; on a mono track use your host's mono-to-stereo routing. Linux is on the roadmap.
- Processing
- On your machine, on the CPU The models are embedded in the plug-in. No audio leaves your computer, and no connection is required once the plug-in is activated.
Hear it on a real stem
Hear the difference
Press play, then switch between Original and Cleaned. The same moment plays in each.
0:00/0:00
Ay, Corazón · htdemucs vocal · Leak Clean
Preparing audio…
Try it on your own stems14-day DAW trial · no card
Hear what was removed & keyboard controls
Removed plays at the original's gain. If you can hear words or melody in it, the setting is too strong.
Click this comparison panel or Tab into it to use shortcuts:
- 1
- Original
- 2
- Cleaned
- 3 or D
- Removed
- Space
- play / pause
- ←/→
- ±5 s
Tip: Removed should sound like drums, instruments and room, not the singer. In the plug-in the same check is the Residual button. Each clip is an htdemucs split of one of five songs, processed by de-leak 0.2.1 with Leak Clean. Choose any of the five songs with the Demo vocal selector above.
Spectrograms are rendered for this demo clip only.
Spectrograms of the Ay, Corazón vocal (vocal stem), rendered from the same Original, Cleaned and Removed files the player streams, on one shared level scale. Log frequency 150 Hz–20 kHz, time left to right, brighter is louder. Click or drag to move the playhead.
Watch the plug-in in action
Two songs, each one 10-second vocal excerpt through every engine, then what each control does. Turn sound on.
Recorded with de-leak AU 0.2.1: the editor was captured while the audio rendered, frames mapped to audio time. Engines are level matched; the removed signal plays at the original's gain. In both clips the Gate is set gentler than the shipped Medium preset. These clips do not demonstrate live latency. English and Korean captions available.
New to de-leak? Start with one vocal stem.
Watch the 9-minute walkthrough: what a stem splitter leaves in a vocal, the two engines, Residual, then Gap Close and Strength on five songs. Jump to any chapter.
Insert it on the vocal, leave Medium loaded, play, and turn on Residual. You should hear drums, instruments and room, not the singer. If you hear the singer, step down to Default; if bleed survives, try Leak Clean.
Narrated, 8 minutes 56 seconds, de-leak 0.2.1 AU on five songs. The Gate in the demos is set gentler than the shipped Medium preset; the “more is not always better” chapter compares the two. English and Korean captions.
Earlier version
Browse individual video tutorials and comparisons
An earlier demo, recorded in May 2026 with a 0.1 build before the Leak Clean engine: de-leak-rt and de-artifact on the same drum bus.
What does de-leak-rt do?
de-leak takes the bleed a stem separator leaves in a vocal — drums, synths or other instruments still audible under the voice — out of the stem, and the Residual monitor plays only what it removed so you can judge every decision by ear. Version 0.2 is built and tuned for vocal stems from Demucs, Suno and Udio splits; the Instrument and Master presets remain for other stems, which are harder to clean than vocals. It is for mixing and bouncing, not live monitoring. VST3, CLAP and AU on macOS, plus AAX for Pro Tools on Apple Silicon; VST3 and CLAP on Windows; 14-day free trial.
Is this the right tool?
de-leak removes content that does not belong in a stem. It does not separate stems, and it does not repair the texture of what does belong. Everything below is what that covers and what it leaves to other tools.
Reach for it when
- A vocal you split with Demucs or htdemucs still carries the drums or the synth under it.
- Suno or Udio gave you stems, and the vocal stem has the band leaking through.
- You want to hear exactly what a cleanup removed before you commit it to a bounce.
- You mix in Logic, Live, Reaper, Bitwig, Studio One, FL Studio or Pro Tools on Apple Silicon.
Reach for something else when
- The problem is the texture of the material itself — metallic shimmer, watery smear, codec hash. That is de-artifact. The two are frequently used on the same source.
- You are chasing one ringing resonance rather than bleed. That is ENZYME.
- You still need to split the mix. Separate it first; de-leak cleans the stems afterwards.
- You need it while tracking or performing live. The reported delay is about a quarter of a second.
What it will not do
- Separate stems
- It processes the stem you give it. Splitting a mix into parts is a separator's job; taking out what the separator left behind is what de-leak is for.
- Remove bleed that sits on top of the voice
- Where an instrument plays at the same time and frequency as the singer, removing it takes some of the voice too. Residual tells you when that is happening: if words or melody come through it, back off.
- Clean every stem as well as a vocal
- 0.2 is tuned for vocal stems. Guitar, synth and other broadband stems are harder to clean without damage; the Instrument and Master presets stay conservative for that reason.
- Run everywhere
- VST3, CLAP and Audio Unit on macOS 11+; VST3 and CLAP on Windows 10 x64. AAX on Apple Silicon Macs only, with both engines. No standalone application and no Linux build.
How it works
de-leak looks at the stem's spectrum frame by frame and decides, band by band, how much of each frame is the voice and how much is leak. The header always shows which engine is shaping the sound.
- GateA learned 16-band leak gate: one gain per band per frame, with factory presets and per-band depth. Fast and preset-driven.
- Leak CleanA 64-band direct-sound model that keeps the voice and lets the rest go. Runs on the vocal alone, with the original mix on a sidechain, or after the Gate.
- ListenResidual plays dry minus wet — exactly what was removed — and the Removed map shows which bands were cut over the last 12 seconds.
Per frame. Gate → Clean runs the two lanes in series.
Features
One fixed, reported latency
About 257 ms at 44.1 kHz in the Standard setting (11,334 samples), fixed for the life of the instance and reported to the host, so delay compensation keeps it aligned. Offline bounces and playback sound the same. The Latency setting is read when the plug-in loads, so a change applies on reload.
Residual and Bypass
Residual plays dry minus wet — listen to exactly what is being removed; if you hear the vocal, it is too strong. Bypass is latency-aligned, so the A/B stays sample-accurate.
Three bands on the spectrum
Drag the two crossovers, drag a band's handle to set how deep it is cut, Shift-drag for extra reduction, and solo low, mid or high. With Leak Clean the handles become per-band caps: no cut deeper than that line.
The Removed map
A scrolling map of the last 12 seconds shows which bands were cut and when. Since 0.2.2 rows where the input is silent are drawn dimmed instead of looking like the strongest removal.
Bright, off by default
An optional HF Restore stage adds model-generated content above 8 kHz to the cleaned signal. It is synthesized, not a recovery of the original, and moving it never changes the reported latency.
8 factory presets
Pick from the ◀ name ▶ bar in the header. A preset writes its values into the controls, and they stay editable; the name gains a * once you move away from it. Save your own with Save As.
| Preset | Use it for | Character |
|---|---|---|
| Medium | Any stem, unless you have a reason to pick another | The tuned starting point between removal and damage. On a vocal, compare it with Default. |
| Default | Any stem, when you want a gentler first pass | Catches obvious leak and little else. |
| Vocal | Raw vocal stems from Suno and Udio | Keeps the formants and works the high band harder. |
| Instrument | Drums, bass, guitar, synth | Conservative and transient-safe; subtle leak may stay. |
| Master | A full mix or the master bus | Almost transparent; only egregious leak is touched. |
| AI Ghost | Suno Other and Percussion stems with ghosting above about 4 kHz | Aimed at the air band. Cymbal sizzle and hi-hat air go down with it — judge by ear. |
| AI Heavy | Coarse AI vocals | Strong, non-linear reduction. More removal, more damage. |
| Aggressive | Hearing the leak on its own with Residual | Maximum reduction — for inspection, not for the mix. |
Try it on your own stems
de-leak runs in your DAW, so the trial is the plug-in itself: every engine, preset and control for 14 days, no card. There is no browser version; for texture cleanup in the browser there is de-artifact's free preview.
macOS 11+ (Universal) or Windows 10 x64 · VST3, CLAP, AU, AAX on Apple Silicon · stereo
- InstallDownload the installer below and rescan plug-ins. The trial starts the first time the plug-in opens online.
- InsertPut de-leak on the separated vocal. Medium is loaded; delay compensation lines the track up.
- ListenTurn on Residual. Drums and room are what you want to hear; the singer means back off.
- ChooseTry Clean and Gate → Clean on the same passage and keep the one that sounds right.
Try it free for 14 days, buy when you're sold
14-day full trial — every engine, every preset, every parameter, no gaps. After 14 days the plug-in inserts 50 ms of silence every 10 seconds until you activate.
macOS — Apple Silicon + Intel · Windows — x86_64 · Linux — roadmap
Not sure yet whether your track was AI-generated? Run it through the free AI music detector — no account, no upload of the master.
System requirements
de-leak runs its models inside the plug-in on the CPU, so a single instance is comfortable on every supported configuration.
- macOS
- Universal Binary — Apple Silicon (arm64) + Intel (x86_64) in one .pkg · macOS 11 (Big Sur) or later
- Windows
- Windows 10 x64 or later · .exe installer or portable .zip · VST3 + CLAP
- CPU
- Minimum Apple M1 / Intel Core i5 (2-core, 2.0 GHz+) / AMD Ryzen 5 · Recommended Apple M1 or later (multi-instance)
- RAM
- Minimum 8 GB · Recommended 16 GB (large stem sessions, multi-instance)
- Audio buffer
- Minimum ≥ 128 samples · Recommended 256 samples for everyday use
- Sample rate
- Host rates from 8 to 384 kHz; the model runs at 44.1 kHz inside the plug-in.
- Channels
- Stereo, processed per channel, with an optional stereo or mono sidechain for Leak Clean.
- Hosts (macOS)
- VST3, CLAP and AU (Apple Silicon and Intel) — Logic Pro, Live, Reaper, Bitwig, Studio One · AAX: native Pro Tools on Apple Silicon (Gate and Leak Clean; Intel and Rosetta Pro Tools are not supported)
- Hosts (Windows)
- VST3 + CLAP — Live, Reaper, Bitwig, Studio One, FL Studio
- Linux
- Roadmap.
FAQ
- Does de-leak separate a full mix into stems?
- No. de-leak cleans a stem you have already separated; it does not split a full mix into vocals, drums and other instruments. Use a stem separator first, then insert de-leak on the vocal stem to reduce the accompaniment left under the singer.
- Does Leak Clean need the original mix or a sidechain?
- No. Leak Clean can run on the vocal stem alone. If your DAW routes the original mix to its sidechain input, it can use that reference while the sidechain carries signal. In native Pro Tools on Apple Silicon, the AAX build uses the key input.
- Can de-leak remove all bleed without changing the voice?
- No. Dense overlaps, backing vocals and reverb can make wanted and unwanted sound difficult to separate. Residual monitoring lets you hear what is removed, including any vocal detail. Reduce the processing if words, breaths or the melody become prominent in Residual, and compare in the full mix.
- Does de-leak work in Logic Pro and Pro Tools?
- Yes. Since 0.2.1 the macOS installer includes an Audio Unit for Logic Pro and an AAX build for native Pro Tools on Apple Silicon (Gate and Leak Clean; Intel and Rosetta Pro Tools are not supported). Other hosts use VST3 or CLAP on macOS and Windows.
- Gate, Clean or Gate → Clean — which one?
- Start with the Gate and the Medium preset. If bleed survives, switch to Clean, which runs a separate model on the vocal alone, or to Gate → Clean, which runs it after the Gate. Turn on Residual for each and keep the one where the removed signal has the least of the singer in it.
- An offline bounce comes out silent or short.
- The host needs plug-in delay compensation enabled. de-leak reports a fixed latency to the host — about 257 ms at 44.1 kHz in the Standard setting. Most modern DAWs handle this automatically; allow a render tail, too.
- Will it work on a non-44.1 kHz session?
- Yes. The plug-in resamples host rates from 8 to 384 kHz to the model's 44.1 kHz and back. Nothing to configure.
- How does it differ from a regular noise gate?
- A noise gate opens and closes on level. de-leak decides on content, band by band and frame by frame: it stays open for quiet singing and closes on loud bleed, which a level gate cannot tell apart.
- What if my problem is separation residue rather than bleed?
- They are different defects and want different treatment. Bleed is another instrument present in the wrong stem — what this plug-in targets. Residue is the separator's own reconstruction damage smeared through the stem. Assess them separately instead of raising intensity to suppress an instrument: the bleed-removal workflow covers the bleed case, and cleaning a separated vocal stem separates the two by ear and fixes each.
- I upgraded from 0.1.x and my Logic automation moved.
- In 0.2.0 four parameters were removed and Audio Unit parameter IDs follow the parameter order, so AU automation lanes saved with 0.1.x may now drive a different parameter. Check and reassign de-leak automation in those projects. VST3 and CLAP are not affected.
- Is there a free trial?
- Yes — 14 days of full quality with no gaps, every engine, preset and parameter, no card. The same installer becomes the licensed build the moment you enter a key. After the trial the plug-in inserts 50 ms of silence every 10 seconds until you activate.
Further reading
Try it on your next stem session
One license, all formats. Free updates within v1.x.