The code, model weights, and test sets are public.
📄 Paper: https://t.co/9ENWNBl1bX
💻 Code: https://t.co/qrUSCiKrEO
🤗 Model: https://t.co/VTAlUSR1sB
Joint work with @pswoboda3 at HHU Düsseldorf. See you at NeurIPS!
Our paper got into #NeurIPS2026! 🎉
"Transcoda: End-to-End Zero-Shot Optical Music Recognition via Data-Centric Synthetic Training"
Transcoda reads sheet music and writes the notes as text. It has 59M parameters, trains in 6 hours on one GPU, and beats a model 16x its size. 🧵
On rendered test pages, Transcoda gets 18.46% error (OMR-NED, lower is better). Legato, a 943M model, gets 43.91%.
Real scans come next. The camera-ready adds 206 real pages we transcribed by hand, plus an updated model. Both come out in about two weeks.
Setup takes 15 minutes.
Two free Anki add-ons. Audiovisual Feedback (231569866) and Auto Rate Typed Answer (2142833407).
Any deck with word, reading, meaning, and sentence fields works. Core 2k/6k is a great place to start.
So I am changing how I make content. Not just piano. Language learning, gym, whatever I am actually doing at the time. No more "the best way to do X". Just what I did and what the outcome was.
Here is the full video
https://t.co/bExcb7eLDX