evidoria

← Back to browse

Good practice Imported

NorHand — Norway's National Library Opens Its Handwriting-Recognition AI to the Whole Sector

Norway · Oslo · See the Norway profile · See the Oslo profile

Evidence: Descriptive / self-reported Top 49% 60/100 · Ask Evidence Copilot about this practice

Norway's National Library built NorHand, an open AI handwriting-recognition model trained on ~400 historic hands (4.0% character error rate), which has helped digitise about a quarter of its scanned handwritten collection and been adopted by other Norwegian institutions.

~400 hands
Historic hands used to train the NorHand model
4.0 %
Character error rate (CER) of NorHand
~25 %
Share of scanned handwritten collection processed and published
6 billion parameters
Parameters in the NB-GPT-J-6B model

Details

Maturity
Established
Promoter
Nasjonalbiblioteket (National Library of Norway)
Period
2023-2026
Keywords
cultural heritage, digitisation, public libraries, NLP

Context

Norway's National Library (Nasjonalbiblioteket) has pursued full digitisation of its holdings since 2005, and by the mid-2020s had digitised almost all of its books and over half of its newspapers. Its AI-Lab has released a family of openly licensed Norwegian-language models, including NB-BERT, the 6-billion-parameter NB-GPT-J-6B, and the NB-Whisper speech models.

Objectives

Build an open handwritten-text-recognition (HTR) model to digitise historic handwritten manuscripts and make them full-text searchable.

Activities

In September 2023 the Library published NorHand, an HTR model built on the Transkribus platform and trained on roughly 400 different hands, including manuscripts by authors such as Sigrid Undset. Both the model and its training data were released openly.

Results

NorHand reaches a character error rate (CER) of about 4.0%, and by release had already helped process and publish around a quarter of the Library's scanned handwritten documents as full-text-searchable items. The Library reports that several other Norwegian institutions have since adopted it for their own handwriting-recognition needs.

Conclusions

These figures are self-reported by the Library and its tooling partner rather than independently audited, and no external benchmark study or academic evaluation of NorHand's real-world accuracy across institutions was found; the claim that "many" other institutions have adopted it is not itself quantified.

Implementation
Indicative cost
Low (< €50k)
Time to results
Medium (1–3 years)
Staffing & skills
National Library of Norway AI-Lab researchers and engineers

Conditions for success

  • an existing large digitised text/audio corpus to train and validate on
  • partnership with an established HTR platform (Transkribus)
  • open licensing of both model and training data to enable sector-wide reuse

Common failure modes

  • self-reported accuracy figures without independent audit or academic benchmark
  • scale of adoption by other institutions is not quantified

Where it fits

Governance type
national library / cultural heritage institution
Scale
national, with cross-institutional reuse
Income level
high-income

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Replication kit

Reusable artefacts from this practice — as published by their sources.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Similar practices you may find useful