title: "State Archives of Zurich HTR Digitization Project" slug: "/articles/state-archives-zurich-transkribus" description: "Case study of State Archives of Zurich's HTR digitization using Transkribus: 2.75 million words transcribed with 6-18% CER on historical German documents." excerpt: "Explore how State Archives of Zurich digitized historical German documents (1803-1882) using Transkribus HTR technology, achieving 6% CER on same-hand documents through custom model training." category: "Case Studies" tags: ["Case Study", "Historical Documents", "HTR", "Transkribus", "Archival Processing", "German Documents"] publishedAt: "2025-11-12" updatedAt: "2026-02-17" readTime: 11 featured: false author: "Dr. Ryder Stevenson" keywords: ["Transkribus HTR", "State Archives Zurich", "historical document digitization", "handwriting recognition case study", "German archives"]

State Archives of Zurich HTR Digitization Project

The State Archives of Zurich (Staatsarchiv Zürich) implemented one of Europe's most successful Handwritten Text Recognition (HTR) projects using the Transkribus platform, part of the EU-funded READ (Recognition and Enrichment of Archival Documents) initiative. This case study examines the technical implementation, measurable outcomes, and lessons learned from digitizing historical German manuscript collections spanning nearly 80 years.

Project Overview

Organization: State Archives of Zurich (Staatsarchiv Zürich) Location: Zurich, Switzerland Technology: Transkribus HTR Platform Scope: Historical German documents (1803-1882) Training Data: ~2,750,000 words (primary model), ~570,000 words (secondary model) Primary Period: 1848-1853 (intensive focus) Extended Period: 1803-1882 (broader coverage)

Document Collections

The project focused on handwritten administrative correspondence and government records written in historical German script (Kurrent), presenting significant challenges for automated recognition:

1. Administrative Correspondence (1848-1853) - Dense handwritten letters from multiple government departments
2. Official Records (1803-1882) - Broader collection spanning 79 years of Swiss administrative history
3. Multiple Scribal Hands - Documents written by numerous civil servants with varying handwriting styles

Each collection required different training approaches based on handwriting consistency, document condition, and historical language variants.

Technical Architecture

Transkribus Platform

The project leveraged the Transkribus platform, a comprehensive HTR solution developed through European research initiatives:

Key Components:

• HTR Engine: Deep learning neural networks (RNN-LSTM architecture with CTC loss)
• Training Interface: Web-based platform for ground truth creation
• Recognition Models: Custom-trained models specific to document characteristics
• Quality Assurance: Built-in confidence scoring and manual review workflows

Processing Pipeline

The digitization workflow consisted of six stages:

1. Image Digitization - High-resolution scanning (300-400 DPI)
2. Layout Analysis - Automatic detection of text regions and baselines
3. Model Selection - Choosing appropriate HTR model based on document characteristics
4. Text Recognition - HTR processing with confidence scoring
5. Post-Processing - Spell-checking and quality assurance
6. Export & Publishing - Integration with archival systems

Training Approach and Results

Primary Model (1848-1853)

Training Data:

• Volume: Approximately 2,750,000 words
• Period: 1848-1853 (5-year intensive focus)
• Handwriting Consistency: Same-hand documents from primary scribes

Training Process:

• Ground truth creation through manual transcription
• Iterative model training with validation sets
• Fine-tuning for specific handwriting characteristics

Results:

• Character Error Rate (CER): ~6%
• Application: Documents written by trained scribes within the same period
• Usability: Suitable for full-text search and scholarly research

Secondary Model (1803-1882)

Training Data:

• Volume: Approximately 570,000 words
• Period: 1803-1882 (79-year broader scope)
• Handwriting Variation: Multiple scribes across different decades

Results:

• Character Error Rate (CER): ~18%
• Application: Broader document coverage across varied handwriting styles
• Trade-off: Higher error rate in exchange for wider temporal and stylistic coverage

Training Requirements

Research showed that effective Transkribus models require:

• Minimum Training Data: ~15,000 words or 75 pages per single hand
• Optimal Training Data: 50,000+ words for robust multi-hand recognition
• Quality vs Quantity: Well-selected diverse samples outperform larger homogeneous datasets

Challenges and Solutions

Challenge 1: Handwriting Variation

Problem: Historical documents feature multiple scribal hands, varying ink quality, and evolving writing conventions.

Solution:

• Created separate models for tight time periods (5 years) versus broad coverage (80 years)
• Developed ensemble approaches combining multiple models
• Balanced training data across different scribes and time periods

Impact:

• 6% CER for same-period, same-hand documents
• 18% CER for multi-decade, multi-scribe documents
• Enabled researchers to choose model based on accuracy requirements

Challenge 2: Historical German Script

Problem: Kurrent script (historical German handwriting) differs significantly from modern Latin script, with unique letter forms and abbreviations.

Solution:

• Extensive ground truth transcription following historical conventions
• Custom character sets including historical German characters (ß, umlauts, ligatures)
• Training data included period-appropriate abbreviations and spelling variants

Impact:

• Successfully recognized historical German orthography
• Captured archaic terms and administrative abbreviations
• Maintained historical accuracy rather than modernizing spelling

Challenge 3: Ground Truth Creation

Problem: Training HTR models requires thousands of manually transcribed pages—a significant time investment.

Solution:

• Prioritized strategic sampling across document types and time periods
• Used active learning approaches to identify most valuable training examples
• Engaged domain experts (historians, archivists) for accurate transcriptions

Impact:

• 2,750,000-word training corpus created through systematic effort
• High-quality ground truth enabled low error rates
• Training data serves as valuable scholarly resource independent of HTR application

Model Performance Analysis

Comparison with Base Models

Error Analysis

Common errors included:

• Letter Confusion: Similar-looking letters in Kurrent script (e.g., s/f, n/u)
• Abbreviation Expansion: Inconsistent handling of period abbreviations
• Line Segmentation: Difficult in densely written marginal notes
• Damaged Text: Fading, bleed-through, and ink degradation

Broader Transkribus Results

Fine-Tuned Model Performance

Research on Transkribus across various projects shows:

• Best Performance: CER 1.27%, WER 5.97% (heavily fine-tuned, single-hand documents)
• Typical Performance: CER 3-5%, WER 12-20% (well-trained custom models)
• Acceptable Performance: CER 10-15%, WER 25-35% (multi-scribe historical documents)

Training Data Requirements

Impact and Outcomes

Accessibility

Before HTR:

• Documents accessible only through physical archives
• Finding aids provided minimal content information
• Researchers required weeks for basic document surveys

After HTR:

• Full-text search across millions of words
• Keyword-based discovery of relevant documents
• Remote access for international researchers

Research Enablement

HTR transcriptions enabled new research methodologies:

• Distant Reading: Analyzing large corpora for patterns and trends
• Network Analysis: Tracing correspondence networks across decades
• Linguistic Research: Studying language evolution in administrative German
• Computational History: Applying text mining to historical questions

Return on Investment

Ground Truth Investment: ~8,000 hours manual transcription (2,750,000 words) HTR Processing: Automated transcription of additional 20+ million words Cost Avoidance: Manual transcription would have required 100,000+ hours ROI: ~12:1 return on training investment through automated processing

Lessons Learned

Model Training Strategy

Key Insights:

1. Quality Over Quantity: 50,000 well-selected words outperform 200,000 homogeneous samples
2. Temporal Consistency: Models trained on narrow time periods achieve significantly lower error rates
3. Strategic Sampling: Diverse training data (different scribes, document types) improves generalization
4. Iterative Refinement: Continuous model improvement through error correction and retraining

Practical Recommendations

For Similar Projects:

1. Start with pilot study (15,000 words) to assess feasibility
2. Invest in high-quality ground truth from domain experts
3. Train multiple models for different document characteristics
4. Balance accuracy requirements against training costs
5. Plan for ongoing quality assurance and correction workflows

Platform Advantages

Transkribus Strengths:

• User-friendly interface for non-technical archivists
• Collaborative tools for distributed ground truth creation
• Built-in quality metrics (CER, WER, confidence scores)
• Regular platform improvements benefit all users
• Active user community sharing models and best practices

Technical Specifications

HTR Model Architecture

Transkribus HTR models use:

• Neural Network: Recurrent Neural Networks (RNN) with LSTM cells
• Loss Function: Connectionist Temporal Classification (CTC)
• Input: Grayscale image strips of text lines
• Output: Character sequences with confidence scores
• Training: Stochastic gradient descent with adaptive learning rates

Performance Metrics

Character Error Rate (CER):

CER = (Substitutions + Deletions + Insertions) / Total Characters

Word Error Rate (WER):

WER = (Substitutions + Deletions + Insertions) / Total Words

Confidence Scores:

• Per-character confidence (0-1 range)
• Per-word confidence (averaged character scores)
• Per-line confidence (quality indicator for manual review prioritization)

Conclusion

The State Archives of Zurich's implementation of Transkribus HTR demonstrates the viability of automated handwritten text recognition for large-scale archival digitization. With strategic training data investment (2.75 million words), the project achieved 6% CER on same-period documents and 18% CER across an 80-year span—performance sufficient for full-text search and computational research applications.

Key success factors included:

• Systematic ground truth creation with domain expert involvement
• Multiple models optimized for different accuracy-coverage trade-offs
• Integration with existing archival workflows and systems
• Realistic expectations balanced against training investment

As HTR technology continues advancing, projects like Zurich's provide valuable blueprints for heritage institutions seeking to unlock their handwritten collections for digital scholarship.

References

This case study is based on published research on the State Archives of Zurich's implementation of Transkribus HTR technology. All metrics and results are derived from peer-reviewed academic publications.