A researcher, twenty-two thousand pages, a tool built for oneself first

Twenty years of research trips to five archives, in Europe and in America: boxes of business correspondence, technical periodicals from the 1920s, photographed hand-held, crookedly, by the thousand. Twenty-two thousand shots, only part of which had been gone through — by hand, page after page, in German.

The tool was built for that fonds. Every rule comes from a lesson paid for on these pages: the reading direction that the camera’s metadata gets wrong one time in three; the manuscript that has to go through an extra-strength reread to limit invented words; the shelfmark you have to go and read at the foot of the page because nobody enters it; the original that must never be recompressed. What works on twenty-two thousand pages should work on yours.

wondarchive is an online service that transcribes, translates and indexes archive photos taken in reading rooms, for historians, doctoral students, archivists and genealogists.

It is hosted in France, at OVHcloud, and the documents deposited there never train any model — that was the condition for putting other researchers’ archives there too, not just one’s own.

On the shoulders of research

Automatic reading of handwriting (known as HTR, for handwritten text recognition) is a research field more than thirty years old. wondarchive would not exist without it.

The first large annotated handwriting databases, such as the IAM database at the University of Bern, made it possible to measure progress. In the late 2000s, Alex Graves, Jürgen Schmidhuber and their colleagues showed that a neural network can read a whole line of handwriting without cutting it into letters: many of today’s engines descend from that idea.

For archives, the European projects tranScriptorium (2013-2015) and then READ (2016-2019), led around the University of Innsbruck, gave rise to Transkribus, run since 2019 by the cooperative READ-COOP. In France, the École pratique des hautes études develops eScriptorium and its engine Kraken, free and open. In Germany, the OCR-D programme equips the libraries that digitise early printed books. HTR-United lists the transcriptions that teams share to train models: without that sharing, most old handwriting would remain unreadable to a machine.

These tools remain the reference for anyone who wants to train their own model or prepare a scholarly edition. wondarchive takes another, more modest path: to read a researcher’s photos quickly, with no settings, so that they can find what they are looking for in their collection.

The name: wondarchive, from wonder and archive. It is sometimes spelled Wonder Archive or Wonderarchive: it is the same service.