In private testing

Take any mix apart.

Drums, bass, vocals, everything else — pulled out of a finished record at studio quality, on your own machine, inside the DAW you already use. A proprietary separation engine written in C++, trained only on audio we actually have the right to train on.

Join the waitlist Watch the demo

AU · VST3 · Standalone — macOS & Windows

Four stems

One record in. Four usable tracks out.

Not a novelty filter — separations clean enough to drop straight into a session, re-balance, re-verb, resample, or rebuild the whole arrangement around.

Drums

Kit, percussion and transients kept intact — attack stays sharp instead of smearing into a wash of cymbals.

Bass

Low end lifted out whole, with the sub kept where it belongs and kick bleed pushed back into the drum stem.

Vocals

Lead and backing voices, with breath, sibilance and tails preserved — not gated into a robot.

Other

Guitars, keys, synths, strings, horns — everything that makes the arrangement, in one coherent bed.

Demo

See it pull a record apart.

A full walkthrough — load a track, separate, solo each stem, and print it back to the timeline.

Demo reel — coming soon

We are cutting the walkthrough now. Join the waitlist and it lands in your inbox the day it goes up.

Technology

Built like an audio tool, not a web upload.

A native C++ engine with a proprietary model at its core. No queue, no upload, no monthly credit balance — the separation happens on the machine in front of you.

Runs on your machine

The engine and the model live inside the plugin. Your audio is never uploaded, never queued on someone else's GPU, never used to train anything.

100% local

Two modes, one engine

A low-latency causal mode for monitoring stems while the transport rolls, and a full-quality offline pass when you need the best result on the print.

Realtime + offline

Native plugin formats

AU, VST3 and a standalone app, built on JUCE. Four stem outputs routed to their own buses so your DAW sees real tracks.

AU · VST3 · Standalone

ARA 2 integration

Separation that happens in the background against the whole timeline, at full quality, instead of re-rendering every time you move the playhead.

On the roadmap

Phase-aware masking

Complex ratio masks instead of magnitude-only, so what comes back out still lines up with what went in. Null-tested against the source.

Proprietary model

Clean provenance

Every dataset, sample library and soundfont behind the shipping model is licensed for it — and the register that proves it is public.

Auditable
Ethics

Most separation models are trained on music nobody licensed.

Ours isn't. That decision cost us months and a measurable amount of quality, and we are not quietly walking it back later.

  • 1

    Register first, download second

    Every dataset, sample library, soundfont and audio source gets a row — source, license, allowed use, proof link, date — before a single byte reaches a disk. No row, no download.

  • 2

    The shipping model trains on cleared audio only

    Synthetic multitrack renders from CC0 and CC-BY instrument libraries, plus vocals from professional singers who consented to their recordings being used this way — and who are credited for it.

  • 3

    Research references never ship

    Public benchmark sets and pretrained models are useful for measuring ourselves. They are marked as reference-only, they never train a shipping model, and nothing is distilled from them — distillation counts as training.

  • 4

    When the license is unclear, we drop it

    Ambiguous terms, viral share-alike clauses, or a maintainer who can't vouch for their own samples: rejected, logged, and not re-litigated because the model would have scored better with it.

Provenance register — excerpt
Synthetic multitrack rendersCC0Trains
Orchestral sample libraryCC0 1.0Trains
GM soundfontMITTrains
Acoustic drum kitsCC0 / CC-BYTrains
Consented vocal corpusCC-BY 4.0Trains
Public benchmark multitracksResearch-onlyNever ships
Pretrained reference modelsNC / protoNever ships
Share-alike drum kitCC-BY-SA 3.0Rejected
Full register published at launch, with proof links per row.
Early access

Be in the first room.

We are letting producers and mixing engineers in a handful at a time, and the people on this list go first. One email when the beta opens — nothing else.

No newsletter, no partners, no tracking. Unsubscribe with one reply.