What they built.
The Trug und Schein project worked with letters written by two people whose handwriting volunteers had already learned to read. In a Transkribus interview, Laura Fahnenbruck explains that earlier manual transcriptions became training material for personalized recognition models. That model-building history is the vendor's account.
The successor archive provides stronger evidence for the result: readers can search and read actual transcriptions. Its own method explains that some letters receive machine-generated drafts followed by collaborative correction. It also documents editorial choices, identifiers and uncertainty markers, making human interpretation visible in the published material.
The original archive and editorial-method page verify real public output. Transkribus's interview supplies model-building details and must remain identified as vendor-reported.
Read the archive's working methodHow the tools work together
Handwriting recognition supplies preliminary text from a scan. Readers check it against the source and preserve relevant spelling and layout features. The archive then associates the reviewed transcription with metadata so people can retrieve a letter by date, writer or topic and understand editorial interventions.
Workflow at a glance
Prepare examples of the handwriting
Previously transcribed letters supplied training material for models suited to the two writers.
Generate and proofread a draft
Recognition produces preliminary text; volunteers compare it with the handwriting and correct it together.
Publish a traceable transcription
Reviewed letters receive identifiers and editorial annotations before publication in the searchable archive.
How you could adapt the idea
Try the pattern on a few letters you are entitled to process and can check with a knowledgeable reader. Keep the original image beside the draft and record uncertain words. Expand only after understanding the error patterns; training a custom model requires reviewed examples, not just a large folder of scans.
Effort, cost and limitations
- Effort
- The project relied on skilled readers, volunteer training and repeated proofreading. Training models was additional work beyond uploading scans.
- Cost
- The project does not disclose a reproducible per-letter or total cost. Account, processing allowances and volunteer effort need separate consideration.
The archive says only part of the collection uses Transkribus; do not describe every published letter as AI-transcribed.
Writer-specific recognition does not establish accuracy on another person's handwriting. The original project has ended in its earlier form; the successor archive preserves the work.
Go to the original work.
Original archive, contributors and transcription methodSearchable published lettersVendor interview with Laura Fahnenbruck, 16 July 2024Shared combined-handwriting model linked by the interviewSource-backed editorial account; PrisAIm did not run or independently benchmark this implementation.