VoxSplit

voxsplit.borck.dev — hear real separation output, then self-host.

Local-first AI stem separation with a modern web UI. Drop in a song, get back vocals, instrumental, or full 4-stem splits — everything runs on your own machine, no accounts, no uploads to the cloud.

VoxSplit is a greenfield web app built on proven open-source separation engines. It is inspired by Ultimate Vocal Remover GUI and stands on the shoulders of that project and its model authors.

Status

Working proof of concept: upload → separate with per-job progress → play/download stems. Models auto-use the best available accelerator: CUDA on Linux/NVIDIA, MPS on Apple Silicon (torch-based models), CPU everywhere.

Models

Weights are fetched on demand (first use of each model) into data/models/ from the community model repos maintained by the UVR and python-audio-separator projects. No manual downloads needed.

Accelerator notes: RoFormer and Demucs run on CUDA/MPS automatically. MDX-Net runs via onnxruntime — CPU on macOS, CUDA on Linux with audio-separator[gpu].

Quickstart

Requires Python 3.10+ and uv (or plain pip — see note below).

git clone https://github.com/michael-borck/voxsplit
cd voxsplit
uv venv
source .venv/bin/activate
uv pip install -e .
voxsplit

Without uv: python3 -m venv .venv && source .venv/bin/activate && pip install -e .

Open http://127.0.0.1:8399, drop a .wav/.mp3/.flac, pick a model, hit Split. The first run downloads the selected model (~60 MB) to data/models/.

Architecture

browser (vanilla JS) ──► FastAPI (backend/main.py)
                            ├── jobs.py     thread-safe job store, 1-worker queue
                            ├── engine.py   python-audio-separator wrapper (lazy import)
                            └── data/       uploads, outputs, model weights

• REST API: POST /api/jobs (upload + model), GET /api/jobs/{id} (status), GET /api/jobs/{id}/files/{name} (stream stem), DELETE /api/jobs/{id}.
• Engines: models run via python-audio-separator, which supports UVR's MDX-Net/VR models, Demucs, and Mel-Band RoFormers.

Credits

This project exists because of the people who built the models and the original app:

• Anjok07 & aufr33 — Ultimate Vocal Remover GUI, the project that made high-quality local stem separation accessible.
• nomadkaraoke — python-audio-separator, the maintained headless engine VoxSplit wraps.
• ZFTurbo — MDX23C weights and Music-Source-Separation-Training.
• Kuielab & Woosung Choi — original MDX-Net AI code.
• tsurumeso — original VR architecture code.
• Alexandre Défossez & Meta (Demucs) — Demucs AI code.
• Bas Curtiz — original UVR logo/icon design.

Deployment (NVIDIA)

On a Linux box with NVIDIA GPUs and Docker + nvidia-container-toolkit:

cd deploy
docker compose up -d --build

Then browse to http://<box-ip>:8399. Model weights persist in deploy/data/.

Roadmap

• Per-job progress reporting (tqdm hooks on torch models)
• Mel-Band RoFormer models (current SOTA)
• Apple Silicon MPS acceleration (automatic for torch models)
• Ensemble mode + post-processing (from UVR's playbook)
• Optional Tauri desktop wrapper

License

MIT — see LICENSE.