VoxSplit

Offline lead/backing vocal separation for macOS, as an FL Studio-compatible VST3/AU plugin, a native SwiftUI app and a command-line tool. All three are thin frontends over one Rust engine that runs ONNX models on the Apple Neural Engine/GPU via CoreML. Nothing leaves your machine, and end users install nothing else: no Python, no runtimes.

Download: https://voxsplit.jeffrey.build (Apple silicon, macOS 14+), or from Terminal:

curl -fsSL https://voxsplit.jeffrey.build/install.sh | bash               # app + plugins + voxsplit command
curl -fsSL https://voxsplit.jeffrey.build/install.sh | bash -s -- --cli   # only the command, no password

song.wav ─► Kim Vocal 2 ─► vocals ─► karaoke model ─► lead_vocals.wav + backing_vocals.wav
                                      Best: BS-RoFormer Karaoke (frazer & becruily)
                                      Fast: UVR-MDX-NET Karaoke 2, 3 refinement passes

Quality	Model	3:10 song on an M2 Max
Best (default)	BS-RoFormer karaoke, Core ML GPU	~2 min (first run ever: about +1 min to compile)
Fast	MDX-Net karaoke, 3 passes	~1 min
Vocal Remover	Kim Vocal 2	~15 s

Two tools, in both the plugin and the app:

• Lead & Backing: the pipeline above. Vocal passes (0–5) sets how many times Kim Vocal 2 runs (0 = the file is already vocals); Refinement passes (1–5) sets how many times the lead/backing model runs.
• Vocal Remover: vocals.wav + instrumental.wav from the vocal stage alone.

Both UIs show the expected time as soon as you choose a song, and a countdown while it runs. Estimates learn from each run on your Mac.

Layout

engine/    Rust CLI (voxsplit / voxsplit-engine): decode → STFT → ONNX Runtime → overlap-add → WAV
plugin/    JUCE 9 plugin (VST3, AU, Standalone): drop zone, Separate, preview, drag-out
macos/     SwiftUI app (SwiftPM): the same flow; VoxSplitKit holds the async Process() runner
models/    ONNX weights (downloaded, gitignored) + committed sidecar JSON hyperparameters
installer/ .pkg pieces: distribution.xml, welcome/conclusion pages, postinstall
site/      voxsplit.jeffrey.build: Cloudflare Worker + static page + install.sh; installers in R2
assets/    the VoxSplit mark (icon.svg → macos/AppIcon.icns via scripts/make-icon.sh)
scripts/   fetch-models, build-engine, build-plugin, build-app, build-pkg, release, embed, sign, test-all
docs/      architecture, bundle layout, engine protocol, measured performance

Build (Apple Silicon, macOS 14+)

Requirements: Xcode command-line tools, CMake ≥ 3.25, rustup. The Rust 1.92 toolchain is pinned in engine/rust-toolchain.toml and installs automatically; JUCE 9.0.4 is fetched by CMake.

scripts/fetch-models.sh     # MDX weights (~120 MB, SHA-256 verified); exports the RoFormer
                            #   via scripts/export-roformer.sh (needs uv; ~1 GB of PyTorch, once)
scripts/build-engine.sh     # → engine/target/release/voxsplit-engine
scripts/build-plugin.sh     # → dist/VoxSplit.{vst3,component}, installed to ~/Library/Audio/Plug-Ins
                            #   (--no-install to skip; quit your DAW first if it has VoxSplit loaded)
scripts/build-app.sh        # → dist/VoxSplit.app
scripts/test-all.sh         # engine unit + CLI tests, headless plugin tests, Swift Testing suite

Bundles are assembled and ad-hoc signed in ~/Library/Caches/VoxSplit/stage and copied to dist/. Building them outside the checkout is deliberate. If the repo lives in an iCloud-synced folder (like ~/Desktop), macOS stamps Finder metadata onto new bundle folders, and codesign refuses those. Run codesign --verify on the staged or installed copies, not on dist/. For distribution, set VOXSPLIT_SIGN_IDENTITY to a Developer ID and notarize.

Release (maintainers):

scripts/build-pkg.sh        # → dist/VoxSplit-<version>.pkg: app + VST3 + AU + `voxsplit`, sharing one
                            #   engine + models in /Library/Application Support/VoxSplit
scripts/release.sh          # build-pkg, then upload to R2, update site/release.json, deploy the site,
                            #   and check the live download (needs `wrangler login`)

Use the engine directly from the checkout:

engine/target/release/voxsplit-engine probe
engine/target/release/voxsplit-engine separate "$HOME/Music/song.wav" --keep-instrumental
cd engine && cargo run --release --example stem_check -- null "$HOME/Music/song.wav" "$HOME/Music/VoxSplit/song"

Using it

FL Studio: rescan plugins, then put VoxSplit on a mixer insert.

1. Drop a song onto the plugin (or click to browse) and press Separate (or Return).
2. When it's done, press ▶ above the stems to hear them all together; each stem's speaker button mutes it. The preview replaces that insert's input while it plays, so you hear it where you're working. Space plays and pauses.
3. Drag a stem row onto the playlist, or press Reveal.

The run keeps going if you close the plugin window, and the result is saved with the project.

App: the same flow in VoxSplit.app. You can also drop songs on its Dock icon or use Open With. Stems are written to ~/Music/VoxSplit/<song>/.

Terminal: readable progress in a terminal, JSON Lines when piped (--format json forces it).

voxsplit separate song.mp3                                   # lead + backing → ~/Music/VoxSplit/song
voxsplit separate song.mp3 --quality fast --passes 3 --reveal
voxsplit separate song.mp3 --tool vocal-remover               # vocals + instrumental
voxsplit play ~/Music/VoxSplit/song --mute backing            # hear the stems together
voxsplit estimate song.mp3

See docs/ for the architecture, bundle layout, engine protocol and measured performance.