Next-Step-OMR

Next-Step-OMR (NSOMR) is an ongoing research project on optical music recognition for real scanned piano music — the degraded, skewed, ink-bleeding scans found on IMSLP, not clean synthetic engravings.

This is a placeholder / project-existence page: no code or weights are released here yet.

What it does

• Input: a single cropped pianoform system (grand staff) from a scanned page
• Output: Humdrum **kern — full symbolic transcription (pitch, rhythm, voicing, ties, key/meter), convertible to MusicXML/MEI
• Built as a full pipeline: page → system detection → crop → transcription → constrained decoding + reranking

Approach (high level)

A vision encoder–decoder transcription model trained on large synthetic corpora, then fine-tuned on human-verified transcriptions of real scanned systems — a growing labeled pool built with a purpose-made annotation/verification workflow — with RL-based post-training on top. Evaluation is preregistered and holdout-disciplined throughout.

Where it stands

On a frozen internal holdout of real scanned IMSLP piano systems (verified ground truth, micro OMR-NED via musicdiff), the current production model is at parity with the strongest public models we have measured on the same crops, with complementary error profiles. Bar-level timing correctness on hard scans remains the open frontier — for every system we have tested, including ours.

Status

Active research, results-first. Code, weights, or labeled data may be released later; nothing here is production-ready. Questions or interest: open a discussion on this repo.

Related

• Project page on GitHub: https://github.com/weselyj/nsomr