Overview of AI models and (AI-powered) OCR/HTR software used for text-processing in humanities research

Commonly used OCR software

AI-powered OCR tools

A [Tesseract OCR workflow], combined with an automated Google translation of the text output, is presented in: Selvaganapathi, G. (2024, July 29). Optical Character Recognition (Ocr) with google translate. Medium. https://medium.com/@gayathri.s.de/optical-character-recognition-ocr-with-google-translate-9c30bfb703d7

OCCAM is integrated with the European crowdsourcing platform Transcribathon, which allows for manually editing transcriptions. The tool was created by Tom Vanallemeersch, CrossLang NV (Belgium), and is offered as a service on https://ai4culture.crosslang.dev/ui/. In the GUI, there are no options to manually improve image segmentation, however, so the tool may only recognise small parts of a scan as actual text. The strength of the tool therefore lies in the evaluation and correction of workflows rather than from-scratch OCR. OCCAM’s support for PageXML correction — using language models, lexica, or validated transcriptions — can help with refining existing OCR outputs in structured formats.

TRANSKRIBUS, especially in its browser-based light version, has an intuitive user interface and is very beginner-friendly, but also much more of a black box. Although the project started as an EU-funded research project, the tool is now under Read Coop, a for-profit business. Therefore, TRANSKRIBUS has introduced a subscription model and gives users who pay priority access to computational resources and better models. Also, they are marketing the tool as an AI-powered, universal platform for text recognition, which raises high user expectations. In reality, community models often underperform on less common data types, such as multilingual materials. Unfortunately, there is also little transparency about the quality of the community models, their training data, or performance measures. As mentioned before, we cannot say anything about their “super models” because we never paid to use them, but the overall promises of “AI” seem exaggerated when the models offered are still trained for rather narrow use cases.

AI models for OCR cleaning and analytical tasks (experimental)

[!CAUTION] The overview of models was created in spring 2025 and we cannot promise to update it regularly, so the information may no longer hold true when you visit the repository. AI development progresses so quickly that it is difficult to keep track, and we are also unable to test all AI models ourselves. This is why you should consult more timely contributions in tech forums for more reliable information on the different models and recommended uses.

Ethical challenges

Working with AI raises questions of research integrity and ethics, especially when historical and multilingual sources are concerned. AI-powered OCR often performs better for widely used languages such as English and Spanish. In addition, AI-based spelling corrections trained on modern and more widespread language use may alter original meanings and ingest anachronisms in texts. AI would need to be trained on more diverse sources (including historical and regional texts) to perform better. At the moment, AI use still requires a high level of human monitoring. When publishing AI-enhanced text, scholars should flag where AI interventions (e.g. spelling corrections) have altered the original and what has been done to check the quality of the AI output. In many cases, direct AI text manipulation is not recommended. But AI can be used to develop controlled vocabularies or create lists of typical errors.