VoxSplit

Local-first AI stem separation. Drop in a song, get studio-grade vocals and instrumentals — on your own hardware, no accounts, no cloud uploads.

MIT licensed CUDA & Apple Silicon SOTA RoFormer models

Self-host in 3 commands View source

Hear it first

Real separation output from VoxSplit. Listen to the original, then the extracted stems — same 30-second clip, no touch-ups.

Loading demo…

Why local-first?

Your music never leaves your machine. No upload queues, no per-minute fees, no expiry links — just a FastAPI service and your GPU doing the work. Batch a whole album while you make coffee.

Run it yourself

git clone https://github.com/michael-borck/voxsplit
cd voxsplit
uv venv && uv pip install -e .
uvicorn backend.main:app --port 8399

Apple Silicon: RoFormer & Demucs models accelerate via MPS automatically. NVIDIA Linux: use deploy/docker-compose.yml for a CUDA build. First model use downloads weights once (~1 GB for RoFormer).

The models

On-demand weights from the UVR / python-audio-separator ecosystem: Mel-Band RoFormer (current SOTA for karaoke/vocals), UVR's MDX-Net line, and Demucs 4-stem — all selected per job from the web UI.