Ultimate Vocal Remover and friends leave Radeon owners on CPU or flaky DirectML. StemSep runs the same model families — BS-Roformer and Demucs — on AMD's official ROCm-on-Windows PyTorch wheels. A 4-stem fine-tuned Demucs pass on an RX 9070 XT takes ~75 seconds; the same job on CPU takes 12–20 minutes.

Synths/pads land in Other on the Demucs presets. The Roformer 2-stem beats Demucs (and UVR's defaults) for clean acapellas.

After a run you get a synced multi-waveform rack: space-bar transport, click-to-seek across all stems, per-stem solo/mute — audition the split before you commit it to a deck or a project. Right-click any stem to save just that piece as a WAV wherever you want it, or reveal it in Explorer — or grab a stem's label and drag it straight out of the app onto a rekordbox deck, an Ableton track, or an Explorer folder.

• Windows 11, AMD RX 9000 / RX 7900-series GPU (ROCm-on-Windows supported)
• AMD Adrenalin driver 26.2.2+ (older drivers → falls back to CPU, much slower)
• uv + Python 3.12, Node 20+, Rust toolchain

npm install
uv sync --project python   # ~2 GB: ROCm PyTorch + audio-separator
npm run tauri dev          # dev mode
npm run tauri build        # installer

First separation per model downloads the model checkpoint (~100–700 MB) into the audio-separator cache.

Stems land next to the source track — drag straight into rekordbox or Ableton:

<track folder>\Stems\<track name>\<preset>\<track name> - Vocals.wav

• Tauri 2 + React 19 UI (src/): drop zone, preset cards, synced multi-waveform stem rack (wavesurfer.js), solo/mute, space-bar transport.
• Rust (src-tauri/): spawns the Python sidecar via uv run, streams JSON-lines progress as Tauri events, kill-on-abort.
• Python sidecar (python/separate.py): audio-separator with ROCm torch pinned by direct wheel URLs in pyproject.toml (so a stray pip install can't silently swap in CPU torch — a known ROCm-on-Windows footgun). AMD's torch wheel requires rocm[libraries], which only exists as an sdist in AMD's wheel directory — a uv dependency override redirects the requirement there.