Automatic transcription of early French manuscripts: which tool, how reliable, what method

Can you automatically transcribe an early French manuscript?

Yes: automatic transcription of a French manuscript, also called HTR (handwritten text recognition), now produces high-quality text on common cursive hands. The rule does not change: you check proper names, dates and abbreviations. Recent AI models read the whole page without prior training on the scribe's hand, whereas classic HTR tools often needed a model adapted to each handwriting.

wondarchive is an online service that transcribes, translates and indexes archive photos taken in the reading room (manuscripts, typescripts, printed material), for historians, doctoral students, archivists and genealogists. A page recognised as handwritten goes through enhanced review, with the strongest model: it reads cursive hands in European languages, French included. The breakdown by type of page is on what wondarchive reads.

Which tool should you choose to transcribe French manuscripts?

To transcribe thousands of photos of French manuscripts taken in the reading room, then read, search and cite them, a ready-to-use service such as wondarchive fits, with no training required. To train a model on one specific hand, Transkribus or eScriptorium are better suited. To decipher a single page, a general-purpose AI assistant can help out. The choice therefore depends on volume, the preparation you are willing to do and how you will use the text.

The detailed comparison of each tool, with its status and where your images go, is on the tool comparison page.

How accurate is automatic transcription of French manuscripts?

Automatic transcription of a French manuscript is of high quality when the photo is sharp and the hand legible; proper names, dates, numbers and abbreviations still need checking against the photo. Reading quality depends first on the photo, then on the handwriting: a letter, report or hand-filled form in a legible hand reads well. Tightly written or heavily abbreviated hands need more review: try them on the 100 free pages before committing a whole collection.

• Reads well: letters, marginal notes, hand-filled forms.
• Always check: proper names, dates, numbers and abbreviations, with a glance at the photo displayed beside the text.
• Nothing is made up: a blank verso or nearly empty page is recognised as such, with no invented text.
• Tolerated: a page rotated by a quarter or half turn is set upright before reading; a tilt of a few degrees is read without loss.

How do you transcribe manuscripts photographed in the reading room, step by step?

With wondarchive, transcribing manuscripts photographed in the reading room takes seven steps, with no file preparation: you photograph, upload the folder, step in only where the machine hesitates, then review, organise, search and export. Transcription runs on the server in batches: you can close the tab, it carries on, and the collection is readable the next day.

1. Photograph flat, sharp, full page: the rules are in the guide photographing in the reading room.
2. Upload the folder for the research trip exactly as the camera wrote it, subfolders included; the original JPEG is kept intact.
3. Straighten: the browser decides each page's orientation; if it hesitates, it shows you the four orientations and you click the right one.
4. Transcribe: each page is read from the full-resolution original; the tool notes the type of document, its language, its date, the dates cited and the proper names.
5. Review in two panes: the photo on the left, the text on the right, original or translated in one click.
6. Organise by place, shelfmark and folio; the shelfmark printed at the foot of the page is read and suggested for the following pages.
7. Search and export: full-text search across the whole collection, export to PDF, DOCX, TEI or CSV.

How do you check and correct a manuscript transcription?

You check a manuscript transcription by first reviewing proper names, dates and numbers against the photo displayed beside the text, then correcting the text directly. In wondarchive, a correction does not erase the machine's reading: its original version is kept and every correction is dated. A collection glossary also guides translation, so that a term is rendered the same way from start to finish.

A quick method: review first the pages you will cite, then search on a name to find its different spellings across the whole collection. How it works is described on the page correcting the transcription.

Can you search for a name across thousands of transcribed manuscript pages?

Yes: once manuscript pages are transcribed, wondarchive offers full-text search across the whole collection. Proper names are indexed with their variants, dates are understood even when imprecise, and you can search for two terms at a distance from each other. Because shelfmarks printed at the foot of the page are read, searching a collection of thousands of shots saves you from reopening them one by one.

Do my manuscript photos stay confidential, and how do I export the text?

Yes: photos uploaded to wondarchive are hosted in France, at OVHcloud, across three data centres, with the originals in Paris. They are never used to train any model, the service complies with the GDPR, and a full export lets you leave at any time. Text can be cited through a stable link and exported by page, by shelfmark or by search, as PDF, DOCX, TEI or CSV. Details are on the page your data.

Follow-up questions

What about a Latin manuscript?

Latin is among the languages transcribed and translated into yours. Printed Latin is read very well; a manuscript goes through enhanced review, with abbreviations and names to check. See the guide transcribing Latin archives.

What about German manuscripts in Kurrent or Sütterlin?

Kurrent and Sütterlin are read like other cursive hands, with enhanced review. The result depends a lot on the hand and the photo: try them on the 100 free pages.

Do I need to train a model on the handwriting in my collection?

No. wondarchive asks for no hand-transcribed ground-truth pages and no training: the reading engine, chosen from a benchmark of about thirty engines, reads the page as it was photographed.

Can I translate a French manuscript for a colleague abroad?

Yes: each page is translated using the end of the previous page and the start of the next as context. Translation is offered in fourteen languages.

How can I try it on my own manuscripts?

A free account gives you 100 pages transcribed, translated and kept for one year, with no credit card. You can judge the reading on your own handwriting before committing a whole collection.

To judge on real material, upload around a hundred of your manuscript photos: open a free account, or first browse the demo.