Opens in a new tab
S U B T R A K T Audio
Menu

Release-ready sound from AI song exports

Adaptive spectral denoise, Tone, stem-aware mastering, and Release Ready loudness, processed offline on your Mac (macOS 14, Apple Silicon).

14-day trial · preview and A/B unlocked · export locked until licensed · then $79 one-time · free 1.x updates. Download and checkout links publish with release (see Pricing).

S U B T R A K T

Stems

Clean

Strength0.90

Tone

Warmth0.55

Master

LUFS−14

Illustrative UI · CSS / SVG faceplate · not a live app capture

The problemAlmost finished. Not release-ready.

  • 01
    Metallic hiss and radio texture

    Broadband grain that spectral denoise pulls down while keeping phase and stereo.

  • 02
    Smeared transients

    Drum hits and consonants lose edge. Stem-aware paths can spare attacks when the separator is available.

  • 03
    Harsh highs

    Vocal sharpness and brittle air. Tone and AI Clean correct the original mix, not a stem re-sum.

  • 04
    Loudness that fails delivery

    Release Ready targets Streaming −14 LUFS / −1.0 dBTP.

Three steps to a cleaner master

STEP 01

Drop a mix or stems

Single Clean for WAV/MP3, or Stem Remixer for a Suno stems folder, ZIP, or local htdemucs separate.

STEP 02

Shape the rack

Flip modules between simple and control-room. Set Clean, Tone, Release Ready, then Preview with A/B.

STEP 03

Export and recall

PCM_24 WAV (optional MP3), HTML report, and a .subtrakt.json sidecar for identical recall.

Four modules. Shared metal. Quiet accents.

Flippable DSP on one graphite chassis. Accents only on stripes, arcs, and LEDs.

Adaptive spectral denoise and AI Clean

Linked-stereo STFT denoise (Bass Hz 350, Strength 0.9). Optional AI Clean uses htdemucs stems as masks on the original mix.

Strength · Bass Hz · Quiet · AI Clean · strength 0…1

Warmth, vocal level, low end

Adaptive body below ~1.5 kHz, vocal level −3…+3 dB, adaptive low-end when the bottom is muddy.

Warmth · Vocal level · Low end

Release Ready and source-aware

Trim, LUFS, 8× true-peak limiter, ~3 ms de-click. Streaming −14 / −1.0 dBTP. Optional source-aware masks.

Loudness · True peak · Clean edges · Source-aware

Stem remix and masks

Rebuild from Suno stems. Core ML FP32 htdemucs. For AI Clean / Tone / source-aware, stems are masks only.

Mixer · Delay · HPF · Mud · Vocal Balance · Limit

Flip a module: front vs technical back

Clean · front

Strength0.90
Bass Hz350
Flip to control room
Clean · control room

n_fft 2048 · hop 512 · Hann
quiet fraction 0.05 · shelf 3 / 5 dB
AI Clean TF masks · linked stereo
reconstruction check −60 dB

Show front

Illustrative UI · front / back flip

Swipe the modules

Scroll-snap carousel of large illustrative modules. Trackpad, touch, and keyboard focus inside the strip.

Strength and AI Clean

Spectral denoise with bass protect. Optional stem-guided AI Clean on the original mix.

Clean

Strength0.90
AI Clean1.0

Warmth, vocal, low end

Three knobs on the original mix. Defaults at zero keep the file byte-identical until you move them.

Tone

Warmth0.55
Vocal+1.0
Low end0.45

Release Ready

Streaming loudness and true-peak ceiling. Clean edges combines trim and de-click.

Master

LUFS−14
dBTP−1.0

Per-stem faders

Mixer gains, mute/solo, and remix polish. Separator stems also feed analysis masks.

Stems

LUFS, TP, GR, correlation

Last-render readouts after Process / Remix. Spectrum and stem levels when available.

Meters

Illustrative UI · CSS scroll-snap · swipe or use arrows

Before / after spectrum

Drag the divider. Curves are illustrative, not a live meter from your file.

BeforeIllustrationAfter

Meters, removed noise, A/B

Preview like tape. Meter like a rack. Spectrum after every render.

  • Pro meters

    Integrated / short-term / momentary LUFS, LRA, true peak, gain reduction, stereo correlation and width, per-stem levels.

  • Listen to removed

    Debug output can write a Removed Noise WAV so you can hear only what was subtracted.

  • A/B comparison

    Tape-machine Preview with A·B and optional level-matched playback. Matching never changes the export.

  • Before / after spectrum

    1/6-octave overlay after a render, with last-process readouts in the Meters module.

Meter bay

Peak

Illustrative UI · analog VU style · not a live meter

HTML processing report

English or Norwegian bokmål. Steps, real parameters, loudness bars, honest AI-provenance note.

Settings sidecar

.subtrakt.json beside the export. Load it later, or drop it on the queue for identical DSP recall.

Offline by default

Processing runs locally. Optional separator models download to Application Support. Anonymous diagnostics are opt-in (default off). See the privacy policy.

Honest about AI provenance

Subtrakt does not hide AI provenance and does not remove watermarks. It is about sound quality and release loudness, not disguising origin.

One-time license. Hear everything in the trial.

macOS on Apple Silicon only (macOS 14+).

Subtrakt 1.x

$59$79

Launch price for about 4–6 weeks, then $79 one-time.

  • Full app during 14-day trial (export locked until licensed) until licensed
  • Free updates within 1.x
  • 2.0 upgrade at $39
  • Standard now · Pro coming later
Buy · $59 launch price

14-day free trial

$0

Preview, play, and A/B. Only export is locked until you buy or the trial ends.

  • No credit card for the trial clock
  • After expiry: open and preview still work
  • Offline Ed25519 license token
Download free trial

Straight answers

What are the system requirements?

macOS 14.0 or later on Apple Silicon. Optional Core ML htdemucs separator (FP32 ~205 MB recommended).

How does the trial work?

14 days. Hear and play everything, including Preview and A/B. Export is locked until a license. After expiry, preview still works.

What about updates?

1.x updates are free. A 2.0 upgrade is planned at $39. Pro may come later.

Refunds?

Refund terms will match the payment provider once checkout is live. Contact support@subtraktaudio.com.

Can I use Subtrakt on commercial releases?

Yes. You remain responsible for rights to AI-generated material and platform AI-disclosure rules.

What does Subtrakt not do?

It does not hide AI provenance, remove watermarks, or run as a real-time DAW plugin in 1.0. Not available on Windows or Intel Macs for the native app.

Is processing private?

Yes by default. Optional anonymous diagnostics never include file names, paths, emails, or license tokens unless you share a 30-second clip. Privacy.