SplitScore

Separate an audio file into 6 stems with BS-RoFormer-SW, then transcribe the stems you choose to MIDI with MuScriptor. A local web console (FastAPI + vanilla JS).

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

Quick start

uvx splitscore

This auto-detects your NVIDIA GPU and installs the correct CUDA torch backend. First run downloads models (~3 GB total) into ~/.cache/.

Requirements

• Windows 11 / Linux, Python 3.13, uv
• NVIDIA GPU recommended — requires CUDA 12.x Toolkit installed for GPU acceleration (auto-detected; falls back to CPU if missing)
• Apple Silicon (MPS) and AMD (ROCm) GPUs are also supported

Setup (development)

uv run python sync.py          # auto-detect GPU, install correct CUDA torch
uv run python -m app           # start server

MuScriptor weights are gated: you need a free Hugging Face account.

1. Open https://huggingface.co/muscriptor/muscriptor and accept the CC BY-NC 4.0 (non-commercial) license.
2. Log in from the terminal:

   uv run hf auth login
   

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

Usage

1. Drop an audio file (wav / mp3 / flac / ogg / m4a / aiff).
2. Separation runs automatically — wait for the 6 stems to appear.
3. Tick the stems you want, adjust per-stem instrument (empty = auto), hit Transcribe selected to MIDI.
4. Download the .mid files. Each stem is named <song>_<stem>.mid.

Every setting (model size, device, precision, temperature, beam/batch size, output folder) is adjustable in the Settings panel and persists across runs.

Notes & limitations

• MuScriptor does not preserve velocity — output is timing, pitch, and instrument.
• Drums and "other" stems tend to transcribe poorly to readable MIDI; melodic stems (vocals, piano, guitar, bass) give the best results.
• MIDI + stem WAVs are written under the configured output folder (./output/<job_id>/stems|midi/).
• One job at a time by design.

License

• Code: MIT — see LICENSE.
• Models: the separation model and the MuScriptor weights are separate and distributed under their own licenses. In particular, 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.