Turn a scanned piano score (an IMSLP PDF, or photos of the pages) into falling notes: watch and hear the piece in the browser, export it as an MP4, or download a MIDI file for Synthesia. Right hand and left hand land on separate tracks, at a tempo you choose.
Synthy runs Audiveris to read the score, then snaps every measure to its time signature before writing MIDI, because optical music recognition gets pitches mostly right on dense Romantic writing but drifts on rhythm. Measures that had to be stretched or squeezed a lot are listed so you know where to check.
Python 3.11 or newer.
pip install -e ".[dev]"Audiveris 5.11 on Linux, without root:
mkdir -p ~/.local/opt && cd ~/.local/opt
curl -sL -o audiveris.deb https://github.com/Audiveris/audiveris/releases/download/5.11.0/Audiveris-5.11.0-ubuntu22.04-x86_64.deb
dpkg-deb -x audiveris.deb extract && mv extract/opt/audiveris audiveris && rm -rf extract audiveris.debOn macOS (Apple silicon; use the x86_64 dmg on Intel):
curl -sL -o /tmp/audiveris.dmg https://github.com/Audiveris/audiveris/releases/download/5.11.0/Audiveris-5.11.0-macosx-arm64.dmg
echo Y | hdiutil attach -nobrowse -readonly /tmp/audiveris.dmg
mkdir -p ~/Applications && cp -R /Volumes/Audiveris/Audiveris.app ~/Applications/
hdiutil detach /Volumes/AudiverisThen the OCR data, on either platform:
mkdir -p ~/.local/share/AudiverisLtd/audiveris/tessdata
curl -sL -o ~/.local/share/AudiverisLtd/audiveris/tessdata/eng.traineddata https://github.com/tesseract-ocr/tessdata/raw/main/eng.traineddataThe Linux build needs a Java 21 runtime on the path; the macOS app bundles
its own. Synthy looks for the binary at ~/.local/opt/audiveris/bin/Audiveris,
~/Applications/Audiveris.app and /Applications/Audiveris.app; set
SYNTHY_AUDIVERIS to use a different install, and TESSDATA_PREFIX for
a different OCR data folder.
Web app:
synthy serveThen open http://127.0.0.1:8000, drop a PDF, set the tempo, and either
play the piece in the browser or download the MIDI. Uploads and
intermediate files live under ~/.cache/synthy/jobs.
The player draws the notes falling onto an 88-key keyboard, right hand green and left hand blue, with piano sound. Space plays and pauses, the arrow keys skip four beats. Sliders set the playback speed, the scale (pixels per beat, so notes can be tall and fast or short and dense) and the note width; settings are remembered. A style menu switches the bar look (matte, bevel, strike edge, capsule), and "Compare styles" shows all four side by side on the same clock. Measure lines and the suspect measure shading can be toggled. When a key is struck again while its note is still held, the earlier bar is cut at the new onset. "Export MP4" renders the same view with audio to an H.264/AAC file, faster than real time, in Chrome or Edge.
Command line:
synthy convert score.pdf -o score.mid --tempo 80 --pages 1-3,5Both print the measures that look suspect: their expected length in beats and what each hand actually read as.
PDF / image -> page PNGs (2550 px wide, grayscale)
-> Audiveris batch export, one MusicXML per page
-> RawScore: notes per staff with raw onsets and durations
-> normalize: scale each hand of each measure to the time signature,
infer a missing time signature, merge ties, flag outliers
-> MIDI: track 1 right hand (channel 0), track 2 left hand (channel 1)
-> player: the same note list as JSON, drawn on a canvas and played
through the FluidR3 piano soundfont; MP4 export encodes
that canvas and an offline audio render with WebCodecs
Measures before the first recognised time signature use the most common measured length (from the shorter staff of each measure, since engine mistakes mostly add duration); a time signature found later is only used from where it appears, because it may be a meter change.
The engine sits behind synthy.engine.Engine, so another recognizer
can be plugged in by writing MusicXML per page.
pytest # unit tests, no Audiveris needed
pytest tests/test_integration.py # runs Audiveris on a real scan if both existDesign notes and the implementation plan are in docs/superpowers/.