Handwriting OCR, benchmarked on IAM
Turn handwritten notes, filled forms, and letters into text. Measured at 1.29% character error rate on the IAM handwriting set — matching frontier vision models, where traditional OCR lands near 49%.
- Reads cursive and print: handwritten notes, filled forms, ledgers, letters
- 1.29% character error rate on the IAM test set — beats the open TrOCR specialist model, on par with frontier VLMs
- Same /v1 parser: page reading order, tables, and bbox citations come with it
- One call: okra upload note.pdf, then okra read <docId>
Frequently asked questions
How accurate is okraPDF on handwriting?
1.29% character error rate (CER) on the IAM test set — the standard handwriting benchmark. That sits in the 'excellent' band (under 2.5%), ahead of the open TrOCR-large specialist model and on par with frontier vision models. Traditional OCR (Tesseract) scores about 49% on the same lines.
Does it read cursive?
Yes. IAM is unconstrained English cursive and print from 600+ writers, and the 1.29% number is measured on exactly that. Messier or historical hands are harder — character error rate climbs as the source degrades.
How is this measured?
120 lines sampled from the IAM-line test split (MIT-licensed, hosted on Hugging Face), scored as character error rate against the ground-truth transcriptions, with punctuation tokenization normalized identically across every system. Full method is on the page.
Related convert from pdf tools