Turn archival scans into a checkable transcript

Goal & fit

Create a corrected, image-linked transcription of a bounded archival sample, with uncertainty and editorial interventions preserved. Searchable text is a working artifact, not proof of an accurate edition.

Suitable when

Collections with adequate images and human expertise to judge a pilot across different hands, layouts and image quality. Recognition can assist a correction workflow when provenance is retained.

Not suitable when

Unlicensed uploads, pages too damaged to support a reading, or a corpus without a qualified checker. Recognition should not invent missing text or silently modernize historical language.

Inputs to gather

• Rights-cleared page images with archive, collection, shelfmark and page identifiers; preserve the original images.
• A transcription policy covering spelling, abbreviations, line breaks, damaged text and uncertain readings.
• Representative pages and a reviewer able to read the relevant script and language; access to a suitable recognition model or support for training.

Suggested workflow

The sequence below is editorial guidance. Numbered references identify sources for tool capabilities, not evidence that this entire workflow was tested.

1. Establish rights and editorial policy

   Check the archive's reproduction and processing conditions, including whether third-party upload is permitted. Assign stable page IDs and document transcription conventions. Keep a diplomatic transcription distinct from any normalized reading or translation.

2. Build a deliberately difficult pilot

   Sample different hands, scripts, columns, marginalia and degraded pages rather than selecting only clean examples. Manually transcribe a reference subset. Reserve some checked pages for evaluation instead of using every reference page to adapt the model.

3. Inspect segmentation before recognition

   Check page regions, lines and reading order. A correct word in the wrong column is still a transcription error. eScriptorium describes segmentation, recognition and correction in one workflow [2]; Transkribus describes recognition and layout tools [1]. Select a route only after checking its suitability for the material.

4. Recognize, then correct against the image

   Run a pilot with a documented model and settings. Review names, dates, amounts, abbreviations and uncertain characters against the scan. Mark illegibility explicitly. Keep model output separate from human corrections so later readers can distinguish a prediction from an editorial decision.

5. Evaluate before extending the run

   Compare held-out pages with the checked reference under the same transcription conventions. If reporting character error rate, record normalization, alignment rules and the denominator. Also inspect omitted lines and reading-order errors, which a single text score can hide. Rework the model or layout when the pilot fails the project's acceptance criteria.

6. Export with provenance intact

   Export text and layout links with page IDs, model details, correction status and unresolved readings. eScriptorium documents ALTO XML and PAGE XML export [2]. Reopen a sample export and confirm that text still maps to the correct image regions before using it for search, extraction or publication.

Alternative tools & routes

Choose by project needs and governance, not by an assumed quality ranking. Current costs, licences and service terms need checking before adoption.

Expected output

A versioned transcript linked to page and line identifiers, the unaltered images, a correction and uncertainty log, and a pilot evaluation note. Keep an expert manual-transcription route available for pages where recognition adds more correction work than value.

Validation checklist

Use these as acceptance checks in your project record. This page does not save completion state.

• Every exported page has the correct shelfmark and image mapping.
• Reading order, marginalia and omitted or duplicated lines were checked, not only spelling.
• High-consequence names, dates and quantities were reviewed against the image.
• Evaluation pages were separated from model-adaptation material, and error reporting states the transcription convention.
• Uncertain readings remain marked and normalized text is distinguishable from diplomatic transcription.

Privacy & research integrity

• Archive access is not automatically permission to upload or redistribute. Historical records may still contain sensitive information about living people or communities.
• Do not let a language model fill damaged passages as if they were transcribed text. Conjectural restoration belongs in a separately attributed scholarly layer.
• A model may perform unevenly across scripts, hands and page conditions. Vendor coverage claims do not establish accuracy on your collection; this guide has not run either platform.

Linked primary sources

Based on stored primary-source extracts accessed on 15 October 2026. They were re-read for these guides, not freshly fetched or independently reproduced. Features and terms may change.

1. Transkribus — official platform overview

   Vendor description of recognition, editing and layout capabilities; performance claims are not independently verified.

   Source accessed: 2026-10-15

2. eScriptorium — About

   Project description of segmentation, human correction, training and ALTO/PAGE export.

   Source accessed: 2026-10-15