Most UK archives with significant handwritten holdings have heard of Transkribus. Fewer have heard of eScriptorium — and fewer still have managed to run it. This article explains what eScriptorium is, why deploying it yourself is harder than it looks, and what a managed eScriptorium hosting arrangement actually provides for UK archives, universities, and research projects.
What is eScriptorium?
eScriptorium is a free, open-source web application for Handwritten Text Recognition (HTR). It was developed at the École Pratique des Hautes Études in Paris and is powered by the Kraken OCR/HTR engine. It allows archivists and researchers to upload images of handwritten documents, train recognition models on specific scripts and hands, and run automatic transcription campaigns across large collections.
It is not a transcription editor in the simple sense. It is a full HTR platform — handling image preprocessing, segmentation (identifying lines on the page), model training, recognition, and export of structured transcription output in formats including ALTO XML, PAGE XML, and plain text.
There is no official central eScriptorium service. The software is provided as open-source code that institutions and partners deploy on their own servers. This is the fundamental difference from Transkribus, which operates as a cloud service with per-page or subscription pricing.
eScriptorium vs Transkribus: the honest comparison
| Feature | Transkribus | eScriptorium (managed) |
|---|---|---|
| Software cost | Subscription / per-page credits | Free (open-source) |
| Hosting | Transkribus cloud (vendor-controlled) | Your data on your/your provider's server |
| Data ownership | Data held on Transkribus servers | Full data sovereignty |
| Model training | Yes — shared and private models | Yes — on your own scripts and hands |
| Export formats | PAGE XML, ALTO, PDF, plain text | PAGE XML, ALTO, plain text |
| IIIF ingest | Limited | Native — ingests IIIF manifests directly |
| GPU requirement | Handled by Transkribus | Required on server for model training |
| Per-page cost at scale | Significant at 10,000+ pages | Zero — no per-page charges |
| Vendor lock-in | Yes — proprietary platform | None — fully open-source |
For small or one-off projects, Transkribus's pay-per-credit model is convenient. For archives with large collections — tens of thousands of pages of parish registers, estate papers, or medieval manuscripts — the per-page costs at Transkribus scale become the dominant argument for eScriptorium. At that volume, a managed eScriptorium instance will pay for itself quickly.
Why self-hosting eScriptorium is harder than it looks
eScriptorium is free software. Installing it, however, is not simple — and running it reliably requires ongoing attention that most archive IT departments are not positioned to provide.
GPU requirements for model training
Training a recognition model on a new hand or script is far more practical with GPU acceleration. CPU training is technically possible but is significantly slower — for anything beyond a small test run, it is not a realistic option. The GPU needs sufficient VRAM for the size of the training job. This is the single biggest practical barrier for archives attempting to run their own eScriptorium instance.
Server configuration complexity
eScriptorium requires a specific server environment: Python, Django, Celery for task queuing, Redis, PostgreSQL, and a web server configuration. Each component needs to be installed, configured, and kept updated. A misconfiguration in the task queue or worker processes can cause recognition jobs to silently fail or queue indefinitely without obvious error messages.
Model management
eScriptorium uses the Kraken engine. Models are trained on ground truth data — annotated transcriptions of sample pages from your collection. Creating ground truth requires manual transcription of a representative sample, which is itself skilled work. Models need to be evaluated, refined, and retrained as you encounter new hands, new scripts, or new document types within a collection.
Storage at scale
High-resolution document images take up significant storage. A collection of 50,000 manuscript pages at archival resolution can easily exceed 500GB. This storage needs to be managed, backed up, and accessible to the eScriptorium server for processing.
What a managed eScriptorium instance provides
When we host eScriptorium for a UK archive, what we are providing is not just a server with the software installed. We provide:
- GPU-accelerated server infrastructure — the hardware required for model training, configured and maintained. You get the benefits of GPU processing without procuring and managing server hardware.
- Installation and configuration — eScriptorium, Kraken, Celery workers, storage, and backup — all set up correctly and tested before handover.
- IIIF integration — if your images are already served via a IIIF image server, eScriptorium can ingest them directly via manifest. This means no duplication of image storage.
- Model training support — guidance on creating ground truth for your specific collection type, training a baseline model, and evaluating recognition accuracy.
- Pipeline integration — transcription output can be routed into AtoM scope notes, TEI documents, or a semantic search index, depending on what your project needs. The output is structured and usable, not just raw text files.
- Ongoing maintenance — software updates, security patching, storage management, and worker monitoring. eScriptorium updates regularly and each release requires careful management of the underlying Kraken models.
Use cases for UK archives
Parish registers
Parish registers — baptisms, marriages, burials — are among the most frequently consulted records in any diocesan or county archive. They are also among the most labour-intensive to transcribe manually. A well-trained HTR model on 18th or 19th-century parish register script can achieve over 90% character-level accuracy on clean images, making it viable as a first-pass transcription for researcher search even without manual correction.
Estate papers and solicitor archives
Legal and estate records from the 18th and 19th centuries — lease books, account books, correspondence — often contain consistent hands trained in legal scripts. Once a model is trained on a specific scrivener or clerk's hand, it can process the bulk of a large archive rapidly.
Chapter acts and institutional records
Cathedral and diocesan chapter acts, institutional minute books, and similar records often run to thousands of pages written in consistent clerical hands. These are strong candidates for HTR — high volume, consistent script, high research value.
University manuscript collections
Medieval manuscripts and early modern documents in university special collections are increasingly being digitised, but transcription remains a bottleneck. eScriptorium supports historical scripts including Caroline minuscule, Gothic book hands, and humanistic cursive, with publicly available pre-trained models that can serve as a starting point before fine-tuning on your specific collection.
20th-century correspondence
Modern cursive handwriting is actually harder for HTR than many historical scripts because it is more variable. However, where a collection contains consistent correspondence from a single author, training a personal hand model is tractable and produces useful results for research access.
The workflow: from scanned pages to searchable text
- Image preparation — images are uploaded to eScriptorium (or ingested via IIIF). The platform handles image preprocessing including binarisation (converting to black and white for better recognition).
- Segmentation — Kraken's segmentation identifies regions and lines on each page. This can be manual, semi-automatic, or automatic depending on the complexity of the page layout.
- Ground truth creation — a sample of pages (typically 50–200 lines for a first model) are manually transcribed within eScriptorium to create training data.
- Model training — the HTR model is trained on the ground truth data. This runs on the GPU and can take hours for a first training run. The resulting model is evaluated against held-out pages.
- Batch recognition — the trained model is applied to the rest of the collection. This is fast — thousands of pages can be processed in hours.
- Export and integration — transcription output is exported in the required format and integrated into AtoM, a TEI document corpus, or a search index.
Realistic accuracy expectations: HTR accuracy depends heavily on image quality, script type, and how much training data is available. For well-preserved 19th-century records in consistent scripts with adequate training data, character error rates of 3–8% (92–97% accuracy) are achievable. For difficult scripts, degraded images, or complex layouts, expect to invest more in ground truth creation and model refinement before reaching useful accuracy levels. We always provide an accuracy assessment before committing to a full processing run.
About the author: Matthew Bruton is a qualified archivist and the founder of Archives Hosting UK. He manages eScriptorium hosting and HTR pipeline work for archives and research projects across the British Isles.