Handwritten Text Recognition Training and Test Set for German Kurrent of the 19th century

German Kurrent Handwritten Text Lines (19th Century) for HTR Model Training

This dataset comprises handwritten manuscripts in 19th-century German Kurrent, prepared for the training of a Handwritten Text Recognition (HTR) model. It contains a total of 9,317 text lines. The data was sourced from the following repositories:

Senatsprotokolle: https://github.com/ubtue/Ground-Truth/tree/main/Senatsprotokolle

Digitale Schriftkunde (Bayerisches Hauptstaatsarchiv):

1806 Reichsstadt Nürnberg, Ratsverläße

1824 Maria Theresia, Privatkorrespondenz

1894 Heroldenamt Akten

Deutsches Textarchiv (DTA):

Libelt, Kosmos-Vorlesungen

Hufeland, Privatbesitz 1829

Erbkam, Tagebuch 1842

Auerbach, Sanders 1869

Auerbach, Sanders 1880

Auerbach, Sanders II 1880

For more details, see the README file of each dataset in the `data/pages/datasetname` folder.

Data Structure

The dataset is organized as follows:

lines/

TestSet/

PNG files: Line images

PageXML files: Transcriptions

TrainingSet/

PNG files: Line images

PageXML files: Transcriptions

ValidationSet/

PNG files: Line images

PageXML files: Transcriptions

pages/

DatasetName/

annotatedJpeg/: full-page images with baselines and text areas visible

pngAndXml/: page images with corresponding PageXML

README.md: dataset-specific metadata and description

Transcription Guidelines

Transcriptions were obtained from the original sources and adapted to follow the OCR-D Level 2 transcription guidelines to the best of the contributor’s knowledge and ability. Disclaimer: I am not a professional linguist and do not read Kurrent fluently. Although care was taken to apply OCR-D Level 2 rules consistently, transcription errors or oversights cannot be fully excluded.

Line Detection

For all datasets except the Senatsprotokolle (which already contained line annotations), line detection was performed automatically using Transkribus, followed by manual correction. Each line was extracted with ascenders and descenders fully included in the text region, while minimizing overlap with adjacent lines.Line Extraction

Line extraction was performed using a Python script, available here.

License

Deutsches Textarchiv data: All content is released under CC BY 4.0.

Bayerische Schriftkunde data:

- Digital reproductions (images): CC0 / Public Domain Mark, per Staatliche Archive Bayerns terms.

- Editorial content and transcriptions: CC BY-NC-SA 4.0