Transkribus

The bottleneck in archival research is reading handwritten historical documents: a collection can only be searched once it has been transcribed, and transcription by humans is slow and requires a high level of skill. Transkribus uses neural sequence models combined with automatic layout analysis and, optionally, language models to convert images of handwritten and printed documents into machine-readable text, and provides a variety of tools for information extraction and recognition from these texts. Users upload document images via a browser, select one of several hundred publicly available models that have been trained on specific scripts, languages, and time periods, and receive transcribed text alongside a confidence score. If no existing model matches the handwriting, the user can transcribe a sample themselves and train a custom model on it. A programming interface enables the batch processing of large collections. The output is searchable, structured text in formats that align with cultural heritage best practice, that renders an archival collection usable as a dataset. The platform is designed for libraries, archives, historians and other researchers, and genealogists who lack expertise in computer vision or language technology, or prefer to use a trusted provider with clear ethical and environmental frameworks to undertake this task.

The platform originated from two EU-funded research projects and has been managed by a European Cooperative Society based in Innsbruck since 2019. Its members include archives, libraries, universities, and personal members committed to working in a cooperative manner. Revenue is generated through the purchase of credits for processing, which now sustains the platform following the end of project funding.

Computation is centralised, with the cooperative owning the full tech stack, including the central model, user software, and hardware. Recognition, model training and information extraction run on graphics processors and servers operated by the cooperative in Austrian data centres which are powered by renewable electricity. These can be accessed via a web browser. Therefore, an institution needs no local hardware, no machine-learning environment to install and maintain, and no technical staff. The barrier to a small archive using neural handwriting recognition is therefore reduced to a credit balance. The corresponding constraint is that documents must leave the institution to be processed, however Transkribus also now offers an “On-Prem” (premises) version for users where this may be an issue. For the general system, processing is confined to servers within the EU, and users retain ownership of their uploads, transcriptions, and any models trained on them. Users also have the ability to delete all three. Users have the responsibility for obtaining permissions, including copyright licensing, for uploading images to the system.

Training a model for a new handwriting style is deliberately within the reach of a single researcher. Roughly fifteen thousand transcribed words — about seventy-five pages — generally suffice for a single handwriting style, and around five thousand words for printed text. Custom models are therefore feasible for individual archival projects. These custom models can then be shared with the Transkribus community to be used in other projects, if they match requirements.

Public guidance covers the technology’s basics, transcription conventions for medieval, early modern and Hebrew scripts, and how to interpret the error metric. There is extensive user documentation and support including training materials. The cooperative offers free access to those teaching with the system and runs a scholarship programme for early career researchers to get free access. There is an active and supportive user community with various fora for exchanging best practice, and helping support the delivery, improvement, and R&D within the tool.

The cooperative reports over half a million users and more than 270 members including institutions and private individuals across over 35 countries, and more than three hundred publicly available models covering a range of scripts, languages and periods, which any user can apply without training their own. Transkribus has now produced accurate transcription of more than 200 million historical texts.

The scale of collection it supports is visible in individual projects: one German state library programme is applying it to the complete historical newspaper holdings of North Rhine-Westphalia, reported at up to twenty million pages. Transcribed collections are published as citable datasets, among them a Swiss cantonal archive, an Alpine association journal and an Ottoman periodical.

Accuracy is reported as character error rate, and the platform publishes guidance explaining what the metric means and how to judge whether a transcription is good enough for a given purpose. Model versioning and the option to retrain support review of what a model produced and when.

The limits of openness are structural. Users own their documents, transcriptions and trained models, and the public model library is shared across the community, but the platform itself is not open source, and running it requires purchased credits on the cooperative’s own infrastructure. The organisation collectively owns everything in the cooperative including the code, the data, and the servers, but these are not available to anyone who is not a member: the business model is to be “as open as possible, as closed as necessary” to protect data assets for the cultural heritage community.

Platform: https://www.transkribus.org/