EXTRAKT

Stem Separation, MIDI Transcription, BPM and Key Detection in one plugin

Anyone who has ever built a bootleg knows the drama. Upload the track to some cloud service, wait, burn credits, end up with four stems that are still not MIDI in your DAW. Then run the whole mix through some Audio to MIDI tool and watch as bass, drums and chords come out as one chaotic note cloud nobody can work with. Something is missing in that workflow. That something is EXTRAKT.

EXTRAKT does four things at once, and all four run on your own machine. Stem separation into Vocals, Drums, Bass and Other based on HTDemucs, the same hybrid transformer model from Meta that cloud services sell by the minute. MIDI transcription per stem via Spotify Basic Pitch, polyphonic, with optional pitch bend for natural vocal modulation. BPM detection via onset energy autocorrelation, accurate to one decimal between 60 and 200 BPM. Key detection via Krumhansl Schmuckler with Camelot Wheel display for DJs.

The actual trick is the order of operations. Standard Audio to MIDI tools throw the whole mix in one pot and try to read notes out of the soup. What comes out is a cloud of bass, chords, snare transients and vocal frequencies that no human can sensibly edit. EXTRAKT separates the mix into four clean stems first, then transcribes each stem on its own. The bassline actually comes out as a bassline. The chords from the Other stem land as clean voicing. The vocal melody as a monophonic line with vibrato. Anyone who has ever tried to build a bootleg in a different key, or replay a bassline harmonically, knows what that is worth.

Everything runs locally. On first launch EXTRAKT pulls the 170 MB models, once. After that, no cloud upload, no credits, no API quota, no waiting on someone else's server chewing through your unreleased material. No recurring cost for stem separation, no recurring cost for MIDI transcription. On Apple Silicon the inference runs through CoreML, so three minutes of audio get separated in 25 to 35 seconds. On Windows and Linux the optimized CPU path with AVX2 takes over.

The workflow is trivial. Drag an audio file into the window, BPM and key appear instantly. Click Separate, wait a moment. Pick a stem tab, click Transcribe, MIDI is there. Export Stems writes four 24-bit WAV files, Export MIDI writes a standard MIDI file with correct tempo map and optional pitch bend events. In parallel, EXTRAKT emits the MIDI data in realtime onto the DAW track so you can route it straight into a software synth. Re harmonization, re arrangement, bootleg, all in one pass.

Three thresholds control transcription per stem: Onset for note starts, Note for frame confidence, Min Length for minimum note duration. Pitch Bend toggles per stem. Leave it on for vocals with vibrato and lead synths with glides, switch it off for bass lines and chord material. Defaults sit per stem type and are remembered between sessions.

For everyone taking tracks apart

Remix culture lives off producers taking other people's tracks apart and putting them back together. It used to mean grabbing a vinyl and cutting it by hand. Then came sampling, then Audio to MIDI, then the cloud services. Each step made something easier and abstracted something else away. Cloud stem splitters were the most recent step. They work well, but they want to see your unreleased material on someone else's servers, they charge per track, and you have to be online.

EXTRAKT brings that step back onto your machine. No account, no credits, no uploads, no internet needed after activation. Your bootleg stays yours. The original idea doesn't become a data point in someone's training set. You're working with the same caliber of tool the big services run in their backend, except it runs locally, as often as you want, without limits.

That's the idea. Four tools, one plugin, local on your machine. That's all it takes to properly take a track apart and put it back together again.

Formats & Systems.

Formats: VST3, AU, Standalone.

Systems:

macOS 11.0+ (Universal Binary, Intel and Apple Silicon).

Windows 10 64-bit or newer.

Linux x86_64 (tested on Ubuntu 22.04+).