Calfa OCR
Dedicated to oriental languages and manuscripts

Use AI-based models to run text recognition and extract data from your scanned documents, archives, books, and push them to their full digital potential.

Handwritten Arabic page in Maghrebi script, digitized in black and white.

Available models

Commercial models
Model Capabilities Usage
Calfa Gen One Armenian HTRArabic HTR In-image location

A generic, cost-effective model for Arabic and Armenian

Try the demo Available as an API
Calfa Gen Extended Armenian HTRArabic HTRLatin HTR In-image locationFields/forms detectionStructured output

A generic model specially enhanced for multilingual documents and special output formats

Try the demo Available as an API
Open models
Model Capabilities Usage
hye-calfa-n OCR model for Armenian Damaged, historical and noisy printed documents
calfa/hye-paddle Open text-recognition models for Classical, Western and Eastern Armenian
calfa/hye-tesseract Open OCR model for Classical, Western and Eastern Armenian Trained on a diverse range of documents: noisy texts, historical fonts, Armenian newspapers

Custom models

We develop tailor-made models for any language or special script, trained on your own documents.

Contact us

Benchmark performance

Measured results on five corporaCharacter recognition rate, volume processed and partners
Corpus Script and documents Volume Recognition rate Partners
The Sīra of Baybars2023–2024 Middle Arabic · 20th-century manuscripts 35,000 pp. 98%average · 96 to 99% depending on the manuscript ANR LiPoL · Ifpo
Catalogue of the San Lazzaro manuscripts2022–2024 Armenian · printed catalogue in 8 volumes, structured as a searchable databaseover 2,000 manuscripts described 6,250 pp. 1.3M words 99.2%of characters recognized, on old and degraded scans of the printed catalogue Mekhitarist Fathers of Venice
TariMa Arabic, Maghrebi scripts · manuscripts, lithographs, prints (17th–20th c.)available in BINA, the BULAC area studies digital library 10,000 pp. 97.2%on average BULAC · CollEx-Persée · DISTAM · IREMAM
Islamic History Open Data Platform Arabic · printed editions of 15th-century historians — 99.5%average · generic model Ghent University, ERC project 2017–2021
ChiKnowPo2024 Chinese · imperial xylographic editionsmodel trained from few annotated pages — 96–99%reading order correct at 97.81% University of Strasbourg · CollEx-Persée

Character recognition rate: share of correctly recognized characters (100% − CER), specific to each corpus.

Use Cases

Page of one of Antoine d'Abbadie's notebooks: list of names in Ethiopic script with their Latin transcription, and handwritten notes in French.

3,000 handwritten pages transcribed by AI

Antoine d'Abbadie's notebooks: several scripts on the same page

Notebooks from his journey in Ethiopia (1837–1848) mixing Ethiopic, Latin, Arabic, Hebrew and Greek scripts, transcribed by a triple HTR model, with the BnF and CNRS–IMAF.

Case study

Manuscript notebook of the Sīra of Baybars open on a table: worn, stained and torn pages in Arabic script.

98% recognition accuracy

The Sīra of Baybars: 35,000 pages transcribed

A specialized HTR model transcribed all 35,000 Middle Arabic pages of the Sīra corpus gathered by the project, copied in the 20th century by several hands, from the most careful to the most cursive, on images of highly variable quality.

Case study See the result (project website)

Double page of an Armenian manuscript digitized from a microfilm, in black and white: two columns of text per page, partly blackened and hard to read.
HTR of microfilmed Armenian manuscripts from the BnF

Recognition directly on microfilms of two 17th- and 18th-century BnF manuscripts, available on Gallica. Damaged documents finally readable.

Have a try on your documents

Get 20 pages offered to test our text recognition in Arabic and Armenian on your documents

Try the demo now

Powerful generic models for text recognition in Arabic and Armenian

To process printed or simple, handwritten documents in Arabic or Armenian script, our ready-to-use general AI models offer you and economical solution for an excellent performance. Try it now online.

Creation of custom AI models able to read the most complex corpora

For any corpora in non-Western language, dialect or from a complex hand : get an unmatched text recognition accuracy by using a custom AI model.