ArchXAI published first versions of AI-powered handwritten text recognition models for Estonian, Latvian and Russian
The ArchXAI project has developed advanced handwritten text recognition (HTR) models for Russian, Estonian, and Latvian, which can be used to convert handwritten historical documents into machine-readable format.
The development of these AI-powered models is based on extensive international collaboration. The ArchXAI project is a collaborative initiative involving the South-Eastern Finland University of Applied Sciences, the National Archives of Estonia, the National Archives of Finland, and the National Archives of Latvia. The ArchXAI project is being implemented with 2,338,265 euros in funding from the European Union’s Central Baltic program.
Large amounts of transcribed and synthetically created training data
To develop the machine learning models, archive experts have transcribed over 13,000 images of archival material in Estonian, Latvian, and Russian from the late 19th and early 20th centuries. A wide variety of historical handwritten materials from all the archives participating in the project have been utilized as training data.
The Russian handwritten text recognition model was also pre-trained on a very large dataset consisting of 138 million synthetically generated lines of Russian text. The final model was further trained on authentic data produced during the project, enabling it to read handwritten text fluently while maintaining a robust language model. Test results show that the models achieve accuracy rates of up to 97–96 percent. The LUMI supercomputer environment, maintained by CSC – IT Center for Science, was utilized in training the models.
The AI-powered models support the accessibility of historical materials
The developed HTR models will be widely used for text recognition of archival materials, which are digitized within the framework of the project. The goal is to make historical documents more easily accessible to researchers and the general public through archives’ online services, while taking into account any restrictions on use.
The models will be further developed during the project by adding new training data, which will further improve recognition accuracy. In addition to the HTR models designed for recognizing handwritten text, OCR models for recognizing typewritten text will also be developed and published for all three languages within the framework of the project.
Demos now available for testing
The source codes for the models have been published on the Hugging Face sites of the National Archives of Finland and Estonia. The models are freely available to anyone who wishes to use them. The models will be updated throughout the project, and both the documentation and model versions on Hugging Face will be updated.
The Estonian and Russian-language models can already be tested via demo services, where users can recognize handwritten text contained in individual images. The demo generates a text version in Estonian or Russian, which can be translated into other languages if desired, for example using AI-based translation tools.
The Russian-language model can be tested on the demo service maintained by the National Archives of Finland: Cyrillic HTR Demo.
You can test the Estonian model on the demo service maintained by the National Archives of Estonia: Estonian Handwritten Text Recognition Demo.
Read about how an HTR model is actually trained: https://centralbaltic.eu/project-news/archxai-project-develops-ai-powered-handwritten-text-recognition-for-19th-and-20th-century-russian/
