Drop a born-digital piano PDF, get a MIDI for Synthesia with the right and left hands on separate tracks.
If any measure cannot be reconstructed with certainty, Legato refuses and tells you where.

How it works · Getting started · Command line · What is supported

Note

The interface is in Brazilian Portuguese. The sample score in the screenshots is an original étude written for the project.

Why

Most PDF-to-MIDI tools run optical music recognition on a picture of the page and guess. Scores exported from notation software do not need guessing: every notehead is a glyph from a music font, with an exact position. Legato reads that vector data directly, solves the rhythm of every measure, and fails closed: a measure that can be read in more than one way blocks the whole MIDI instead of producing a wrong one.

Features

🎼 Reads the vector, not the image	Music-font glyphs and vector paths from mupdf.js. No OCR, no machine learning guesses.
🛑 Refuses instead of guessing	Every measure must add up in exactly one way. A doubtful measure blocks the MIDI and is reported.
✋ Hands and repeats resolved	Hand markings, repeats, 1st/2nd endings, Segno, Coda, D.S. and D.C. are unrolled into playing order.
🎹 Synthesia-ready	Standard MIDI file, PPQ 960, with Right Hand and Left Hand tracks.
🌊 Preview before you download	Waterfall preview with piano playback, per-hand toggles and 0.5× to 1× speed.
🎻 Expressive or exact	Optional phrasing, hairpins, legato, staccato and pedal, while notes stay on the exact grid.

Screenshots

Drop a PDF	Preview and play

How it works

POST /api/convert streams each stage as NDJSON, so the page shows progress while the engine works.

flowchart LR
  A[PDF] --> B[Extract<br><sub>glyphs and vector paths</sub>]
  B --> C[Layout<br><sub>staves, systems, barlines</sub>]
  C --> D[Recognize<br><sub>symbols, rhythm, tuplets, directions</sub>]
  D --> E[Score<br><sub>validate, repeats, hands</sub>]
  E --> F[MIDI<br><sub>SMF, PPQ 960</sub>]
  D -. a measure does not add up .-> X[Refuse and point to it]

Loading

Folder	What it does
src/engine/pdf	Extracts music-font glyphs and vector paths with mupdf.js
src/engine/layout	Finds staves, grand-staff systems and barlines
src/engine/recognize	Symbols, rhythm per measure, tuplets, dynamics and directions
src/engine/score	Builds the score, resolves navigation and assigns hands
src/engine/perform	Optional expressive timeline
src/engine/midi	Writes the Standard MIDI File

Accuracy is measured end to end by a benchmark (bench/): it generates MusicXML as ground truth, engraves it to PDF with MuseScore in several music fonts, converts it back with Legato and compares the notes semantically.

The research behind the approach and the architecture decision are in docs/research/pdf-to-midi-research.md.

Getting started

npm install
cp .env.example .env.local   # DATABASE_URL is optional
npm run dev

Open http://localhost:3000 and press Testar com um exemplo to convert the sample score.

Without DATABASE_URL everything works except the conversion history. With a Neon Postgres URL, Legato stores the metadata, diagnostics and the generated MIDI of each conversion (never the PDF). The table is created on first use, and history is kept per browser.

Command line

npx tsx scripts/convert.ts path/to/score.pdf      # writes fixtures/out/<name>.mid
npx tsx scripts/compare-variants.ts path/to/score.pdf  # expressive vs exact rendering
npx tsx bench/run.ts 20 "Leland,Bravura"          # benchmark, needs MuseScore (mscore) on PATH

What is supported

Grand-staff piano scores exported from Sibelius, MuseScore, Finale, Dorico or Musicnotes: chords, several voices, accidentals, key and time changes, clef changes, ties, triplets and other simple tuplets, grace notes, repeats, 1st/2nd endings, Segno/Coda/D.S./D.C./Fine, tempo marks, rit./a tempo, dynamics, hand markings.

Not yet: scanned PDFs (by design), nested tuplets, tremolos, 8va lines, multi-measure rests, non-piano ensembles. These are rejected with a reason rather than converted wrongly.

Stack

Layer	Technology
App	Next.js 16 (App Router), React 19, TypeScript
PDF	mupdf.js
Playback	Tone.js
Database	Neon Postgres, optional
Deploy	Vercel

License

MIT

_{Built by Matheus Ervilha to learn real pieces in Synthesia.}