A
Kraken recognition model (
.mlmodel) for
OCR of older
Swedish text printed in Fraktur (blackletter). It produces a diplomatic
transcription that keeps historical orthography, the long s (ſ) and period
spelling, and works on scanned page images.
For multi-page PDFs, render pages to images first (e.g. ~300 DPI) and pass them
to Kraken. eScriptorium: import the .mlmodel under Models.
Fine-tuned from
german_print on the Språkbanken corpus
Svensk fraktur
1626–1816 (199 page images with line-level diplomatic transcriptions). The page
images were OCR-bootstrapped and the recognized lines aligned to the ground-truth
line text (≈98 % coverage), producing PageXML for
ketos train. A low learning
rate (
-r 0.0001) was decisive — it let the model improve steadily past epoch 0
instead of drifting away from the strong starting point. See
training/
for the scripts and exact commands.
The validation accuracy is measured on a held-out split of the same corpus and is
therefore optimistic relative to entirely new documents; on a real volume
(1600s Swedish Fraktur) it produced a near-flawless body-text transcription with
preserved long-s and period spelling and no systematic substitution errors.
This is a derivative model. Full chain:
This model is released under
CC-BY-4.0 (see
LICENSE).
