SplitScore

Separate any audio file into 6 stems, then transcribe the ones you want to MIDI. A local web console — no cloud, no accounts after initial model download.

audio ──▶ BS-RoFormer-SW ──▶ 6 stems ──▶ MuScriptor ──▶ per-stem .mid

Table of Contents

• Features
• Installation
• Requirements
• Usage
• GPU Support
• Settings
• API
• Project Structure
• Testing
• License

Features

• 6-stem audio separation via BS-RoFormer-SW (ONNX inference)
• Per-stem MIDI transcription via MuScriptor (small / medium / large models)
• Local web console — drag-and-drop, progress bars, inline audio previews
• Auto-detects NVIDIA, AMD, Intel, and Apple Silicon GPUs
• One-command install via uvx — correct torch + ONNX backends installed automatically
• Configurable beam size, temperature, batch size, and model size
• REST API with SSE progress streaming

Installation

uvx (recommended)

uvx splitscore

Auto-detects your GPU, installs the correct torch variant and ONNX runtime, and opens the app at http://127.0.0.1:8000 (auto-falls back to a free port if 8000 is taken). First run downloads models (~3 GB) to ~/.cache/.

pip

pip install splitscore
splitscore

From source

git clone https://github.com/SeanDolan0/splitscore.git
cd splitscore
uv run python sync.py          # auto-detect GPU, install torch
uv run python -m app           # → http://127.0.0.1:8000 (auto-fallback if busy)
uv run python -m app --port 9000  # explicit port

Hugging Face setup

MuScriptor weights are gated behind a CC BY-NC 4.0 (non-commercial) license. You need a free Hugging Face account before transcription will work:

1. Accept the license at https://huggingface.co/muscriptor/muscriptor
2. Log in from the terminal:

   uv run hf auth login
   

   (or export HF_TOKEN=hf_... in your shell)

Requirements

• Python 3.13 and uv
• Windows 11 or Linux
• NVIDIA GPU recommended
• AMD GPU (Linux, ROCm) works fully for both separation and transcription
• Apple Silicon — separation runs on CPU (ONNX has no Apple GPU provider), transcription uses MPS
• Intel GPU — falls back to CPU for both (no ONNX/ROCm provider available)

Usage

1. Drop an audio file onto the console (wav, mp3, flac, ogg, m4a, aiff).
2. Separation runs automatically — watch the progress bar while 6 stems are extracted.
3. Pick stems — tick the checkboxes for vocals, piano, guitar, bass, drums, or other.
4. Set instruments — each stem gets an instrument dropdown; leave blank for auto-detect, or pick a specific instrument (piano, guitar, sax, synth, choir, etc.).
5. Transcribe — hit "TRANSCRIBE SELECTED TO MIDI" and download the .mid files.

Stems

The separation model outputs 6 stems in a fixed order:

Melodic stems (vocals, piano, guitar, bass) give the cleanest MIDI. Drums and "other" tend to transcribe poorly.

GPU Support

SplitScore auto-detects your hardware on startup:

The sync.py script handles all of this during dev setup. The splitscore CLI handles it automatically for end users.

Settings

All settings are adjustable in the web console's config panel and persist across runs.

API

The app exposes a REST API consumed by the frontend:

Project Structure

splitscore/
├── app/
│   ├── __init__.py        # version
│   ├── __main__.py        # python -m app
│   ├── cli.py             # uvx splitscore entry point
│   ├── gpu.py             # GPU detection + provider resolution
│   ├── main.py            # FastAPI routes, SSE, static mount
│   ├── pipeline.py        # Job orchestration, one-at-a-time lock
│   ├── separator.py       # BS-RoFormer-SW ONNX inference
│   ├── settings.py        # Settings dataclass, JSON persistence
│   ├── transcribe.py      # MuScriptor wrapper
│   └── static/            # Vanilla JS/CSS frontend (no build step)
├── tests/                 # pytest, ~2s, no GPU needed
├── sync.py                # Dev setup: detect GPU, install torch
├── pyproject.toml
└── LICENSE                # MIT

Testing

uv run pytest                              # all tests
uv run pytest tests/test_pipeline.py       # single test file

Tests use fake separator/transcriber factories — no GPU, no Hugging Face token, no real models. The full suite runs in ~2 seconds.

License

• Code: MIT — see LICENSE.
• Models: the separation model and MuScriptor weights are distributed under their own licenses. MuScriptor's weights are CC BY-NC 4.0 (non-commercial) and gated on Hugging Face; that restriction applies to the models regardless of this project's MIT license.