Rastral

What it is

Composers still sketch on paper; the transcription into notation software is the tax. Rastral reads a photograph of a handwritten page and returns notation you can edit and play back. The prototype is deliberately narrow and real: single-line monophonic melodies, real model inference on every upload — no canned results.

The demo exists to answer one question: will musicians actually upload their pages?

How it works

Three parts with one hard boundary between them. A web app handles upload, notation rendering, playback, and the correction flow. A thin gateway forwards images and persists results. A Python ML service does the inference on GPU — the only part that knows how the model works.

web (React)  ──image──▶   gateway   ──HTTP──▶   ML service (GPU)
   render    ◀─MusicXML──  persist  ◀────────   image → { musicxml, midi }
                              │
                              ▼
                          Postgres

Fig. 1 — the contract at the boundary: image in, notation out

Nothing outside the ML service knows how the model works — so the model can be retrained or replaced without touching the product around it.

Honest scope

Handwritten OMR accuracy sits in the tens-of-percent error range — that is the state of the field, and pretending otherwise would be theater. So the product is framed accordingly: a fast draft plus quick correction, not magic transcription. Every result comes from the live model on the uploaded image, and the correction flow is part of the product, not an apology.

Status

Live as a working demo at rastral.com, gauging real demand before deeper investment. Prototype scope: single-line monophonic melodies.