Drop in a song. Get the vocals and the instrumental as two clean WAV files.
No uploads, no account, no subscription. Everything happens on your computer.
- 🔒 Private by design. Your audio never leaves your machine. After a one-time model download, Khopesh works fully offline.
- 🎚️ Studio-grade separation. It runs becruily's Mel-Band RoFormer "deux", one of the best open vocal-separation models, on your graphics card.
- ⚡ Fast. A 3½-minute song splits in about 45 seconds on an Apple M5 MacBook Pro.
- 🎞️ Takes almost anything. MP3, WAV, FLAC, M4A/AAC, AIFF, and the audio track of video files (MP4, MOV).
- 📚 Batch friendly. Drop a whole album. Khopesh works through the queue one song at a time.
- 🛟 Never overwrites. Stems are saved next to the original. Split the same song again and you get
Song (2) - …instead of losing your first take. - 🪶 Gentle on your computer. Memory-friendly defaults keep the rest of your system usable while it works.
Grab the latest build from the Releases page.
| Your computer | Download | Notes |
|---|---|---|
| Mac with Apple Silicon (M1 or newer), macOS 13+ | …mac-arm64.zip |
Signed and notarized. Unzip and drag Khopesh to Applications. |
| Windows 10 or 11, 64-bit | …win-x64.zip |
Uses any modern graphics card through Vulkan, falls back to the CPU. |
| Windows with an NVIDIA card and the CUDA 12 toolkit | …win-x64-nvidia.zip |
Adds the CUDA engine. Without the CUDA 12 runtime installed it quietly uses Vulkan instead. |
Note
Windows: the app isn't code-signed yet, so SmartScreen may say "Windows protected your PC." Click More info → Run anyway. Khopesh also needs the Microsoft Visual C++ Redistributable (x64). Most PCs already have it, and Khopesh tells you if yours doesn't.
 |
 |
 |
| 1. Download the model Once, 238 MB. Checked with SHA-256. |
2. Drop your songs Or click to choose. Several at once is fine. |
3. Grab your stems Saved next to each song. |
For every song you get two files in the same folder:
Midnight Drive.mp3
Midnight Drive - VOCALS.wav ← just the voice
Midnight Drive - INSTRUMENTS.wav ← everything except the voice
flowchart LR
song["🎵 Your song<br/><sub>MP3 · WAV · FLAC · M4A · MP4 · MOV</sub>"]
prep["audio-prep<br/><sub>decode · downmix · resample to 44.1 kHz stereo</sub>"]
model["Mel-Band RoFormer “deux”<br/><sub>ggml on Metal · Vulkan · CUDA · CPU</sub>"]
vocals["🎤 Song - VOCALS.wav"]
inst["🎸 Song - INSTRUMENTS.wav"]
song --> prep --> model
model --> vocals
model --> inst
Khopesh is a small Electron app wrapped around two native command-line tools:
audio-prep(C++, innative/audio-prep) turns whatever you drop in into the exact format the model expects. miniaudio handles WAV, MP3 and FLAC. Everything else goes through your operating system's own decoders (AVFoundation on macOS, Media Foundation on Windows), so there's no ffmpeg to install. Resampling uses r8brain.bs_roformer-clifrom BSRoformer.cpp runs the separation model with ggml. On a Mac it uses the GPU through Metal, including the newer Metal tensor path. On Windows Khopesh picks the fastest engine that loads: CUDA, then Vulkan, then CPU.
The model listens to the song in chunks of a few seconds at a time. That's what keeps memory use predictable.
Open them with the wrench in the top-right corner. The defaults are the sweet spot for most computers.
| Setting | Options | What it does |
|---|---|---|
| Chunk size | Small (default), Medium, Large | How much audio the model processes at once. Uses about 1.9 GB, 3.8 GB or 7.5 GB of graphics memory. In listening tests Small sounded no different from the larger sizes, and on a 16 GB Mac Large can freeze the whole system. |
| Overlap | 1 (default) to 4 | How many times each part of the song is processed and blended. Higher values can smooth rough spots: 2 takes about twice as long, 4 about four times. |
Khopesh has no telemetry, no accounts, and no analytics. The only time it touches the network is the one-time model download from Hugging Face, which is verified against a SHA-256 checksum before it's used.
Can I use the stems commercially?
The Khopesh app is MIT-licensed. The separation model, Mel-Band RoFormer "deux" by becruily, is licensed CC BY-NC 4.0, which only allows non-commercial use. For commercial work, check with the model's author. And of course, the song itself belongs to whoever owns it.
My computer felt sluggish while splitting.
Separation uses your graphics card hard. Keep Chunk size on Small, and close other heavy apps for long queues. Khopesh keeps your computer awake while a queue runs, so you can walk away.
It says “Unsupported file type.”
Khopesh accepts MP3, WAV, FLAC, M4A, AAC, AIF/AIFF, CAF, MP4, MOV and M4V. Convert anything else to one of those first.
Windows says the audio engine couldn't start.
Install the Microsoft Visual C++ Redistributable (x64) and try again.
Intel Mac? Linux?
Not yet. On Macs Khopesh needs Apple Silicon. The code is cross-platform, so a Linux build mostly needs someone to package it. Pull requests welcome.
You'll need Node.js 20+, Git and CMake 3.20+, plus Xcode Command Line Tools on macOS or Visual Studio 2022 Build Tools on Windows.
git clone https://github.com/aalmonzer/khopesh.git
cd khopesh
npm install
npm run setup # builds audio-prep + the separator, downloads the Windows engines
npm run dev # opens Khopesh with hot reload| Command | What it does |
|---|---|
npm run setup |
Builds or downloads the native tools into resources/bin. Sources and downloads are pinned and checksum-verified. |
npm run dev |
Runs the app in development mode. |
npm test |
End-to-end test of the real app: first-run download, a queue of songs, settings and cancel. Set KHOPESH_MODEL=/path/to/becruily_deux-Q8_0.gguf to skip the download. |
npm run screenshots |
Runs the same test and regenerates every README image, the demo GIF and the app icon. |
npm run art |
Re-renders just the icon, banner and hero image from the existing screenshots. |
npm run package:mac |
Builds the macOS zip. Signs with your Developer ID if one is in your keychain; notarizes when APPLE_ID, APPLE_APP_SPECIFIC_PASSWORD and APPLE_TEAM_ID are set. |
npm run package:win |
Builds both Windows zips. From a Mac, run npm run setup -- --windows first (needs brew install mingw-w64). |
Pull requests and v* tags run the GitHub Actions workflow, which builds every download and keeps them as workflow artifacts.
Project layout
src/
main/ Electron main process: window, queue IPC, model download, runs the tools
preload/ The small, typed bridge the interface is allowed to use
renderer/ React interface
shared/ Types shared by all three
native/
audio-prep/ C++ decoder + resampler (miniaudio, r8brain, AVFoundation, Media Foundation)
scripts/
setup.mjs Builds or downloads native tools
render-art.cjs Renders the icon and README artwork
tests/
e2e.mjs Drives the real app with Playwright
- becruily for the Mel-Band RoFormer "deux" model, and chenmozhijin for BSRoformer.cpp and the GGUF conversions.
- ggml, miniaudio and r8brain for the heavy lifting.
- The name: a khopesh is the curved bronze blade of ancient Egypt. It cuts one song cleanly in two.
Full license details are in THIRD_PARTY_NOTICES.md.
MIT © Khopesh contributors. The separation model is downloaded separately and carries its own license, CC BY-NC 4.0.