de-leak-rt

Real-time leak gate for Demucs / htdemucs stem separations. ML classifier detects inter-stem bleeding per-frame; multiband gain removes it without touching the music. F1 = 0.993 on MUSDB18-HQ.

$149$99one-time Launch price ends Jul 31 Use as API →

FormatsVST3 · CLAP · Standalone
Latency23 ms (low) / 46 ms / offline
Speed134× realtime on M1
Min. specmacOS 11 · M1 / Intel · 8 GB

de-leak-rt plug-in interface — 3-band leak gate spectrum view with low / mid / high frequency bands, monitor toggles, model selector, and gate parameters

7 presets, tuned per stem type

Pick by what kind of stem you're cleaning. Each preset is validated against a real corpus of Suno, Udio, and Demucs separations. Fine-tune later with the gate sliders.

Preset	Use it for	What it does
Default (Universal) first try	Any stem when you're not sure of the source	Conservative — gain ≈ 0.94, mid protected
Vocal	Suno / Udio raw vocal stems	Preserves formant; high band aggressive on RVQ
Instrument	Drums, bass, guitar, synth — transient-heavy	False positives minimized; transients protected
Master	Full mix or master bus	Almost transparent — gain ≈ 0.98
AI Ghost new	Suno Other / Percussion stems with 4 kHz+ ghosting	Air band cut −3.6 to −4.4 dB
AI Heavy	Coarse RVQ vocal	Gate shape 2 + extra GR 6 dB
Aggressive	Diff-mode auditioning, leak-only solo	Maximum reduction — for inspection, not mixing

Three ML models, three character profiles

All preserve vocals; they differ in how aggressively they handle reverb tails and how much non-vocal content they let through.

Model	Trained on	Best for
Default (large_v3)	MUSDB stem leak	Conservative — real vocals always safe
Balanced (v7) recommended	+ 859 HDF5 reverb / clean / FMA neg	Reverb suppression with vocal preservation
Aggressive Reverb (v5)	+ reverb-aware augmentation	Hard reverb tails — risk of vocal damage

Built for stem work

3 latency variants

Low-latency 23 ms for tracking. Balanced 46 ms for everyday use. Offline ±174 ms (bidirectional context) for reference parity.

134× realtime

4-minute mono input renders in ~1.8 s on M1. The classifier is light enough to run on the audio thread — no worker thread, no priming delay.

Multiband solo

Toggle low / mid / high bands independently to localize ghost noise. Solo + Diff together = audition the leak in just one band.

Wet / Dry / Diff modes

Diff outputs dry − wet — listen to exactly what's being removed. Boost the headphone +30 dB; if you hear the vocal, you're over-gating.

HF preservation

Multiband gain with frequency-aware crossover. The gate only applies below the crossover; above it you decide how much HF to protect (0 → 1.0).

Bit-exact Python parity

+137 dB SDR vs. the PyTorch reference. The plug-in is the reference implementation — no quality compromise from the Rust port.

How it works

de-leak-rt frames each STFT bin, classifies it as content vs. leak, and applies multiband gain. Causal — no lookahead — so live monitoring works.

audio (44.1k mono) → STFT → mag → LeakGate (ONNX) → gate_t (T,)
                  → multiband gain (per-freq exponent w/ hf_comp)
                  → spec * gain → ISTFT → clean

The classifier is a 95 K-parameter MLP (~370 KB ONNX) — it makes one decision per STFT frame. Multiband gain is applied with HF compensation so the high-band content you want to preserve isn't pulled down with the leak you don't.

System requirements

de-leak-rt's classifier is a 95 K-parameter MLP (~370 KB ONNX) — light enough to run on the audio thread with no worker. Even an entry-level M1 hits 134× realtime, so a single instance is comfortable on every supported configuration.

Requirement	Minimum	Recommended
CPU	Apple M1 / Intel Core i5 (2-core, 2.0 GHz+)	Apple M1 or later (multi-instance)
RAM	8 GB	16 GB (large stem sessions, multi-instance)
Disk space	~5 MB (ONNX embedded into the bundle)	—
macOS	11 (Big Sur)	13 (Ventura) or later
Audio buffer	≥ 128 samples	256 samples for everyday use
Sample rate	Any — rubato resampler 8 k–384 k. Internal model runs at 44.1 kHz.
Channels	Mono / stereo (per-channel processing).
Hosts	VST3 + CLAP — Live, Reaper, Bitwig, Studio One · Standalone (Core Audio)
Windows / Linux	Roadmap. macOS first.

FAQ

Why do I get clicks when toggling the band solo?: Fixed — solo transitions use a one-pole ramp (~30 ms half-life). If you still hear them, please report with the host and SR.
Diff + solo together — what should I hear?: Only the dry-minus-wet of the active band. If a vocal appears in the residual you're over-gating that band; back off intensity or raise HF compensation.
Offline bounce comes out silent or short.: The host needs PDC enabled. de-leak-rt reports latency = host_chunk_samples (~93 ms @ 44.1 k). Most modern DAWs handle this automatically.
Will it work on a non-44.1 kHz session?: Yes — rubato resamples 8 k to 384 k transparently. The model always runs at its native 44.1 k internally.
How does it differ from a regular noise gate?: A noise gate triggers on RMS below threshold. de-leak-rt's classifier was trained to distinguish musical content from leak per-frame — it stays open during sub-threshold content and closes during loud leak.

Try free, then download

14-day full-feature trial — no card required. Email us, get a license key in your inbox, and download the plug-in once it's live.

Try free for 14 days

Email license key

One trial per email per plug-in. The license key arrives via email and works on a single machine.

Download for macOS

Coming soon

Beta builds ship over email while we finalize Developer ID notarization. Submit a trial above and you'll be on the early-access list.

de-leak-rt-X.Y.Z.pkg Coming soon

macOS — Apple Silicon + Intel · Windows / Linux — roadmap