DMSI 2026 Kalamazoo: Introduction to Handwritten Text Recognition

Full-day workshop at the International Medieval Congress, Western Michigan University, 13 May 2026.

Introduction to Handwritten Text Recognition

  • Event — Digital Medievalist Summer Institute (DMSI) at the International Medieval Congress
  • Location — Sangren Hall 1720, Western Michigan University, Kalamazoo
  • Date — Wednesday, 13 May 2026
  • Sessions — 09:45–11:45 and 14:15–16:15

Session 1 ↓ Session 2 ↓ All Teaching Materials

Overview

This workshop offers a hands-on introduction to Automated Text Recognition (ATR) for medieval and early modern manuscripts. The morning session builds a shared conceptual foundation — what the technology is, where it stands today, and how to use it — while the afternoon gives participants direct practice on their own material and widens the view to include modern command-line and AI-based approaches.

No prior technical experience is assumed. Participants are encouraged to bring their own scanned manuscript images for the afternoon session.

Link to the slides

Session 1 — Morning

09:45–11:45 · Sangren Hall 1720

Time Duration Topic Resources
09:45 25 min ATR: Where are we at the moment? Introduction
10:10 15 min Round of introductions: ATR — where are you? —
10:25 15 min Key concepts: data, ground truth, training, epistemological decisions Introduction → pipeline
10:40 15 min Integrated Transcription Environments (ITEs) eScriptorium
10:55 25 min How to use eScriptorium eScriptorium → workflow
11:20 5 min eScriptorium vs. Transkribus —
11:25 5 min Business models, open science, and HTR-United HTR-United

ATR: Where are we at the moment?

A concise state-of-the-field overview covering:

  • The distinction between OCR (printed text) and HTR (handwriting) and why it matters for medieval studies
  • Current performance benchmarks — what CER figures look like in practice for different script types and conditions
  • The two dominant platforms (eScriptorium/Kraken and Transkribus) and the landscape of community-trained models
  • The shift since 2024 toward large, transferable base models (CATMuS Medieval, TRIDIS v2) that the whole field can build on

→ Background reading: Introduction to ATR, Open-Source Models

Round of Introductions: ATR — where are you?

A structured round in which each participant briefly shares:

  • What kind of material they work with (script, language, century)
  • What they have already tried (if anything) with ATR
  • What they hope to get out of the day

This shapes the afternoon’s hands-on session to the actual needs of the group.

Key Concepts

Four foundational ideas that underpin every HTR project:

Data
Ground truth is the fuel of any HTR system. Quality matters more than quantity: 50 carefully transcribed pages outperform 500 inconsistent ones.
Ground Truth
The manually verified, authoritative transcription of a set of document images. Defines what “correct” means for a given model.
Training
The process of adjusting a model’s parameters so that its output matches the ground truth. For most humanities projects this means fine-tuning an existing base model on project-specific material, not training from scratch.
Epistemological decisions
Every transcription encodes choices — what counts as a character, whether abbreviations are expanded, how diacritics are handled, what Unicode codepoints represent which glyphs. These are not technical details but scholarly decisions that shape what the resulting texts can and cannot be used for.

→ Background reading: Introduction → Transcription Policies

Integrated Transcription Environments (ITEs)

What is an ITE?
A web-based platform that supports the full HTR workflow in a graphical interface: image import, layout annotation, transcription, model training, and export. ITEs (Integrated Transcription Environments) make HTR accessible without requiring command-line expertise.
eScriptorium
Open-source (MIT), self-hostable, built on the Kraken engine. All models and data remain under the researcher’s control. The platform of choice for open-science and reproducibility-focused projects. → eScriptorium section
Transkribus
Commercial (freemium), cloud-hosted, proprietary engine (PyLaia). Large pre-existing model library and a polished interface. Credits-based pricing for production transcription.

How to Use eScriptorium

A live demonstration and guided hands-on covering:

  1. Import — uploading images or an IIIF manifest
  2. Segmentation — running the default segmentation model; correcting baselines in the graphical editor
  3. Transcription — applying an existing model; editing transcriptions in the line panel
  4. Training — starting a fine-tuning run from corrected ground truth
  5. Export — downloading ALTO XML, PAGE XML, or plain text

Participants follow along on the workshop instance at escriptorium.fr (account registration required in advance).

→ Detailed guide: eScriptorium → The eScriptorium Workflow

eScriptorium vs. Transkribus

eScriptorium Transkribus
License Open source (MIT) Proprietary
Hosting Self-hosted or institutional Cloud
Engine Kraken PyLaia
Models Open, exportable Platform-locked
Cost Free (infra cost only) Free tier + credits
Best for Reproducibility, open models Rapid start, large model library

Business Models, Open Science, and HTR-United

A brief framing of why platform choice is a scholarly decision as well as a technical one:

  • Data sovereignty: who owns your ground truth and your trained model?
  • Reproducibility: can another researcher exactly reproduce your transcription pipeline in five years?
  • Community infrastructure: HTR-United as the shared catalogue and standard for open HTR data and models

→ HTR-United section

Session 2 — Afternoon

14:15–16:15 · Sangren Hall 1720

Time Duration Topic Resources
14:15 30 min Test your own material eScriptorium
14:45 20 min Show your results —
15:05 20 min Introduction to VLMs and command-line tools VLMs, TrOCR
15:25 20 min Developing workflows and identifying crucial decisions —
15:45 30 min Questions and inputs —

Test Your Own Material

Participants apply what they learned in the morning to their own scanned images in eScriptorium:

  • Upload 2–5 pages of their own manuscript or archival material
  • Run segmentation and check/correct the detected baselines
  • Apply an appropriate pre-trained model (recommended starting points below)
  • Review the output and estimate CER visually

Recommended starting models by material type:

Material Recommended model
Latin or French medieval manuscript CATMuS Medieval 1.6.0
Documentary / charters TRIDIS v2
Medieval Hebrew BiblIA family
Greek minuscule Meleagre / GreekHTR
Middle High German (Gothic) Inzigkofen models

Show Your Results

A brief show-and-tell: participants share one page of their transcription output with the group. Discussion points:

  • What worked well? Where did the model struggle?
  • What kind of errors appear most frequently — OCR noise, abbreviation confusion, unusual letterforms?
  • What would need to happen to improve the results — more ground truth, a different base model, manual correction?

Introduction to VLMs and Command-Line Tools

A conceptual and practical introduction to the tools beyond ITEs:

Vision Language Models (VLMs)
Large multimodal models (GPT-4o, Gemini, Llama) that can transcribe text from images in a zero-shot prompt. Flexible but inconsistent; best for exploration or low-resource scripts with no training data available. → VLMs section
TrOCR
Microsoft’s transformer-based OCR model, available via Hugging Face Transformers. Requires Python scripting to use but integrates well with NLP pipelines. → TrOCR section
Kraken command line
All eScriptorium actions can be performed programmatically via the Kraken CLI, enabling batch processing, scripted training, and integration into larger digital workflows. 
# Apply a model to a folder of images
for img in pages/*.jpg; do
  kraken -i "$img" "${img%.jpg}.txt" ocr -m catmus-medieval.mlmodel
done

Developing Workflows and Identifying Crucial Decisions

A structured group discussion on what a realistic ATR workflow looks like for participants’ own projects. Key questions:

  1. How much ground truth do I need? — Rule of thumb: 20–50 pages for a usable fine-tune; 100+ for a strong domain model.
  2. Which transcription policy? — Decide before you start annotating; changing it later invalidates all existing ground truth.
  3. Fine-tune or train from scratch? — Almost always fine-tune. A strong base model (CATMuS, TRIDIS) dramatically reduces the data requirement.
  4. How do I validate quality? — Hold out 10% of pages as a test set; report CER on that set, not on the training data.
  5. Where do I share my model and data? — Zenodo for archival DOIs; HTR-United for discoverability; eScriptorium for immediate reuse within a project.

Questions and Inputs

Open floor for questions, follow-up discussion, and practical next steps. Topics that often arise:

  • How to get access to eScriptorium for my institution
  • How to find collaborators or existing ground truth for my script
  • How to cite ATR-produced transcriptions in a scholarly publication
  • What to do when no pre-trained model exists for my material

Preparation for Participants

To get the most out of the workshop:

  • Create an account at escriptorium.fr before the session
  • Bring 5–10 scanned pages of your own manuscript material (JPEG or PNG, ideally 300+ DPI)
  • Read in advance (optional but helpful): Introduction to ATR and the eScriptorium section
  • Take the Quiz to check your understanding beforehand: Quiz
Back to top