OCR and HTR for your printed and handwritten documents

Un texte exploitable, vérifié par des humains.

We process both the files we digitise and the ones you already have: books, typewritten cards, magazines, newspaper titles or handwritten registers.

For manuscripts, recognition models are trained on your own collections from transcriptions reviewed and validated by our teams: accuracy improves as the project goes on.

What we handle

• Books and cards: Printed and typewritten text.
• Newspapers and magazines: Complex layouts, columns, articles.
• Manuscripts: Latin, Arabic or old cursive scripts.
• Duplicated documents and imitation handwriting: Stencils and cursive fonts imitating handwriting.
• Your existing files: OCR and HTR of images already digitised.

Technical specifications

• Printed text: Multilingual OCR, over 600 languages
• Manuscripts: HTR trained on your collections, over 75 languages
• Delivered files: Text PDF, PDF/A, TXT, XML ALTO, TEI XML
• Qualité : OCR : 98 % en automatique, jusqu’à 100 % après relecture ; HTR : 98 % et plus après relecture
• Measurement: Word error rate (WER) and character error rate (CER)
• Quality control: Proofreading and validation of transcriptions, ISO 2859 sampling
• Validation: Formats validated with the CINES tool, SHA-256 checksums
• Uses: Full-text search, analysis, AI

Une méthode, deux voies.

The same six steps, from file received to text delivered, with tools suited to what has to be read: printed matter on one side, manuscripts on the other.

Images are prepared the same way, then each path follows its own settings: engines and dictionaries for printed matter, models trained on your collection for manuscripts. In both cases, an operator proofreads and a sampling check measures quality before delivery.

In both cases

• F : o
• E : m
• R : a

Printed · OCR — Books, cards, press, typescripts

• Intake: TIFF or JPEG images, integrity and completeness check.
• Page analysis: Blocks, columns and reading order, corrected by an operator.
• Model selection: Dictionaries suited to the language and period, historical typefaces included.
• Recognition: Several OCR engines combined by weighted voting.
• Proofreading: Proofreading and correction by an operator.
• Quality control: ISO 2859 sample; any batch outside tolerance is reworked.

Manuscripts · HTR — Latin, Arabic and early cursive scripts, imitation handwriting

• Intake: TIFF or JPEG images, integrity and completeness check.
• Page analysis: Regions, lines and baselines; manual segmentation if needed.
• Model selection: Choice from over 150 public models, by period and script.
• Recognition: Neural networks, private model retrained on your collection.
• Proofreading: Line by line, image on the left and text on the right, guided by word confidence.
• Quality control: Word and character error rates, ISO 2859 sample.

La qualité, à chaque étape.

Quality that is measured, not assumed.

Each batch is measured by word error rate, with character error rate tracked to detect drift. A sample is checked to ISO 2859, and any volume below the threshold is reworked in full.

A performance report gives the measured and validated rate for each batch, with problem areas flagged: illegible entries, missing pages, reproduction anomalies.

Files ready for preservation.

XML ALTO files are validated against the reference schemas, and text PDFs are produced as PDF/A. Formats are checked with the CINES tool, and a SHA-256 checksum is computed for every file delivered to guarantee its integrity.

Deliverables follow BnF recommendations or your platform’s specifications, with a folder structure and file naming that match your rules.

Des textes prêts à l’usage.

Texts ready to join your digital library, your search engine or your analysis projects.

• XML ALTO : Un fichier par page : mise en page, blocs, lignes et position de chaque mot.
• Text PDF and PDF/A: Image with hidden text, at several compression levels, for full-text search.
• TXT and TEI XML: Plain text per image or merged, and TEI encoding for the digital humanities.
• Quality report: Measured rate, problem areas and format validation, for every batch.

Libraries, archives and universities trust us to transcribe their collections.

• Académie des sciences: Digitisation of manuscript documents and HTR processing.
• Government of the Grand Duchy of Luxembourg: OCR processing and XML ALTO structuring of fragile typescripts and manuscripts.
• City of Marseille: Digitisation of archives, press and early documents, with OCR in XML ALTO format.
• Catholic University of Lille: Digitisation of journals and registers, OCR in ALTO format.
• École normale supérieure: Digitisation of scientific and heritage collections, OCR in ALTO format.
• University of Bordeaux and Bordeaux Montaigne: Digitisation and complex OCR, data entry and structuring in XML and CSV.

Frequently asked questions

What is the difference between OCR and HTR?

OCR recognises printed and typewritten text, whose characters are regular. HTR, based on artificial intelligence, reads handwriting, as well as fonts close to handwriting that conventional OCR cannot handle. Both turn your images into text that can be searched, copied and analysed.

What quality can we expect?

For printed matter, OCR reaches 98% automatically, and up to 100% after proofreading. For manuscripts, raw recognition gives around 80% correct words depending on how legible the handwriting is; line-by-line human proofreading brings the result to 98% or more. Quality is measured by word error rate, on a sample following ISO 2859.

How are HTR models adapted to my collection?

We first choose, from over 150 public models, the one that matches the period, language and script. We then train a private model on proofread and validated transcriptions from your collection: the more corrected examples it receives, the more it improves over the course of the project. Over 75 languages are supported, including Arabic.

Which files will I receive?

XML ALTO files (one per page, with the position of every word), text PDFs or PDF/A at several compression levels, plain text per image or merged and, on request, TEI XML, DOCX or CSV tables. Formats are validated and a quality report accompanies every batch.

How does a project run, and how can I estimate its cost?

We process your existing images (TIFF or JPEG) just like the ones we digitise. A trial on a sample validates the model, segmentation and deliverables before production, which is managed by a single point of contact. The cost depends on the number of pages, the type of writing and the quality level required: describe your collection in our quote tool to get an estimate.