On-device AI stem splitter for iOS. Pick any video from your camera roll (or an audio file), split it into vocals / drums / bass / other with Meta's htdemucs model, remix the stems with a mute/solo/pitch/speed mixer, and export the mix, a video with the new mix, or any individual stem as 24-bit WAV — all through the share sheet.
No account. No upload. Nothing leaves your phone.
| Pick from your library | Ready to split |
|---|---|
 |
 |
| On-device separation | Mix, play, export |
 |
 |
- Video-native. The camera roll is the library: pick a video, the audio track is extracted and resampled on device. Audio files import from the Files app too.
- Split from Link. Paste a Spotify track/album/playlist or a YouTube video/playlist: the track is matched on YouTube Music and fetched as AAC (the iOS port of stemsplitter-mac's StemLink, native URLSession instead of the spotdl subprocess), then it splits like any imported file. Link lookups are the only network the app touches.
- 4-stem separation (vocals, drums, bass, other) with htdemucs running on the GPU via Core ML.
- Mixer — mute/solo per stem, mix presets, pitch shift, speed — with live playback.
- Export the mix, any stem as 44.1 kHz / 24-bit WAV, or the original video with the new mix — via the standard share sheet.
- Splitting is fully offline. No analytics, no third-party packages — only Apple frameworks and the bundled model. The optional link fetcher talks to Spotify's embed endpoint and YouTube's player API; playback and splitting never touch the network.
- Streaming engine. Chunks flow through STFT → model → inverse STFT with overlap-add; peak memory stays under 400 MB on a 15-minute split. Instrumental is derived by complement subtraction so stems sum back to the source exactly.
- Xcode with the iOS 18 SDK
- iOS 18+ (device or simulator)
- The converted Core ML model (see below — not included in the repo)
- A free personal Apple Developer team is enough to run on your own device (7-day re-sign); TestFlight needs a paid membership
git clone https://github.com/kristerssaulitis/stemsplitter.git
cd stemsplitter
open StemSplitter.xcodeprojSelect the StemSplitter scheme, pick a destination, and run.
Device: Signing & Capabilities → select your team, then Run. (A free personal team works; the build re-signs for 7 days each time you re-run.)
Simulator: works out of the box but takes the CPU fallback path (the simulator has no GPU-usable Neural Engine path for this model), so splits are much slower than on a real device.
Models/ is gitignored — the app target expects Models/htdemucs.mlpackage at the
repo root (~120 MB, mixed fp16/fp32 precision) and will not build without it.
It is produced by converting Meta's htdemucs weights (MIT licensed) with
coremltools. The Core ML package wraps only the
network core — Core ML has no complex tensors, so the STFT/iSTFT pair around it runs
in Swift with vDSP. The conversion tooling is not in this repo yet; until it lands,
build the mlpackage from htdemucs yourself, or skip the app target and work with the
SwiftPM targets below (they build and test without the model).
If xcode-select points at CommandLineTools on your machine, prefix with
DEVELOPER_DIR:
# Unit tests (model-dependent tests run when Models/htdemucs.mlmodelc exists)
/usr/bin/env DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer swift test
# App build verify (no signing needed)
/usr/bin/env DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer \
xcodebuild -project StemSplitter.xcodeproj -scheme StemSplitter \
-destination 'generic/platform=iOS Simulator' -derivedDataPath .build/dd \
-quiet build CODE_SIGNING_ALLOWED=NO
# Benchmark harness (timings, memory profile, wall-clock linearity)
/usr/bin/env DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer swift run spike <corpus-dir>- Pick. The picker runs out of process — the app never asks for photo-library permission. Videos come from the camera roll; audio files via the Files importer. Sources up to 10 minutes.
- Split. The engine streams chunks with live progress, per-stem peak meters, and a calibrated ETA. Cancel is instant and discards partial output.
- Mix. Stem cards with mute/solo, mix presets, pitch, and speed, over a scrubbing waveform of the source.
- Export. Share the mix, any stem (44.1 kHz / 24-bit WAV, RF64 beyond 4 GB), or the original video re-rendered with your mix — via the share sheet. Export is the only persistence: split outputs live in caches and are purged on next launch.
One-screen single-flow UI (picker → processing → result) on a model-agnostic streaming engine:
StemSplitter.app
UI LAYER (SwiftUI, Sources/StemUI) ENGINE LAYER (Sources/StemCore)
AppFlowView ──state──> ProcessingView AudioExtractor (AVFoundation)
│ │ ├─ AVAssetReader audio-only decode
│ v ├─ AVAudioConverter → 44.1 kHz
PhotosPicker ResultView └─ no-audio / iCloud / offline errors
(out-of-process, ├─ source selector
no library perms) ├─ stem cards + StemEngine (actor, StemEngineProtocol)
│ long-press share ├─ Chunker (size + overlap)
└─ ShareLink WAV ├─ STFT/ISTFT (vDSP) + overlap-add
├─ DemucsEngine (CoreML, GPU→CPU fallback)
SUPPORT (Sources/StemCore/Support) ├─ complement: instrumental := source − vocals
SplitStore Caches/Splits/<uuid>/, └─ WAVWriter (24-bit streaming, RF64 > 4 GB)
purged on next launch
ETACalibrator persisted per-device, PipelineEvent (Contracts/) — the ONE event
keyed by model ID surface engine → SwiftUI (progress, peaks,
ModelStore bundled .mlmodelc, ETA, failure); exactly one terminal event.
preloaded at launch
The StemCore / StemUI / Spike targets and the test suite are defined in
Package.swift (Swift 6, no third-party dependencies); StemSplitter.xcodeproj builds
the app target and compiles the model into the bundle.
Key invariants:
- Streaming pipeline, no whole-file float buffers; peak < 400 MB on a 15-minute split.
- Instrumental is derived by complement subtraction so stems sum to the source exactly.
- Output WAV is 44.1 kHz / 24-bit; mono sources upmix dual-mono; other sample rates pass an explicit resample stage.
- Cancel discards partial output; split WAVs live only in caches — share / Save to Files is the only persistence path.
On-device only: no cloud, no upload, no account, no network calls, no analytics, no
third-party packages. PhotosPicker runs out of process, so the app never requests
photo-library permission. See Release/privacy.html and the
App Store runbook in Release/APP_STORE.md (privacy labels:
"no data collection"; PrivacyInfo.xcprivacy manifest included).
App code: MIT. The htdemucs model weights are MIT licensed by Meta — see the Demucs license.