Kurdish-language OCR and speech research offers practical ways to make historical material searchable. Studies of old publications and Central Kurdish dialogue also show why a scan, a machine transcription and a reviewed edition must remain distinguishable. Preservation and oral-history guidance provide the wider framework.
Preserve the record before processing it
A searchable transcript is an additional way into a collection, not the collection itself. The National Archives of Australia’s Digital Preservation Policy, version 1.4 of 30 June 2020, treats a digital record as content together with linked metadata. Its integrity principles retain the original bitstream, check file integrity and document preservation actions. [1] Applied to a Kurdish newspaper or family recording, this means keeping an unchanged master and a collection identifier while processing working copies. Record the publication or recording date separately from the scan date. A checksum can help detect a changed file; it cannot establish whether the historical statement inside that file is true.
Historical OCR: the Zheen-supported experiment
Blnd Yaseen and Hossein Hassani’s 2024 preprint used digitised publications printed before 1950, supplied by the Zheen Center for Documentation and Research in Sulaymaniyah. They prepared line images with manual transcriptions and adapted an Arabic Tesseract model for the material. The paper reports a 0.755% training error measure, while its separate page evaluation reports 84.02% character accuracy. Those figures describe different evaluations, not interchangeable guarantees. The authors identify spacing, multiple columns and mathematical notation as difficulties. [2] For an archive, the useful lesson is to check the actual page layout and held-out material before accepting extracted names, dates or quotations.
Speech and translation: account for language variety
The 2024 CORDI study by Sina Ahmadi and colleagues transcribed dialogue from Central Kurdish films and series, documenting material associated with Sulaymaniyah, Sanandaj, Erbil, Mahabad, Kalar and Sardasht. Volunteers received guidance and submissions underwent proofreading. The authors tested translation, speech recognition and language identification; speech systems trained on one dataset could perform poorly on another. [3] Film dialogue is not an oral-history interview, and their results do not establish the performance of every current tool. They do support testing the language variety and recording conditions of the collection itself. Translation should remain linked to the original utterance and a reviewer should preserve uncertainty rather than silently replacing an unfamiliar expression.
A useful corpus is still a selected corpus
Geoffrey Haig’s Corpus of Contemporary Kurdish Newspaper Texts contains 483 Northern Kurdish journalistic texts, about 214,000 words, compiled from 1999 online material from Azadiya Welat and CTV. The repository describes the texts as untagged and untranslated, and dates the syntax project to 1999–2001. [4] That makes the collection’s scope inspectable. It does not make it a sample of every Kurdish language variety, historical period or speaking situation. A vocabulary or name found there can guide a search; its presence alone cannot identify a person in an unrelated manuscript. Corpus construction, model training and historical interpretation answer different questions.
Oral history: retain consent, context and the voice
The Oral History Association’s Best Practices call for documented informed consent, opportunities for narrator review where possible, and attention to restrictions on access and distribution. They recommend transcripts or indexes linked by time markers to recordings, preservation planning, and citations to the full interview. [5] In an AI-assisted project, discuss external processing as part of the proposed use rather than assuming that consent to an interview authorises every later transfer. Keep the recording, raw transcript, reviewed transcript and translation connected. A fluent summary may omit hesitation, disagreement or a change of speaker; readers need a route back to the voice and its context.
A small, reviewable pilot for a Kurdish collection
A practical pilot could select ten pages representing clean print, damage, different typefaces and multiple columns. Make a human-reviewed reference, keep machine output separately, and log errors in names, dates and reading order. For audio, select samples with different speakers and background conditions and retain timestamps. These are proposed evaluation steps, not results of a test performed on this site. The newspaper pictured above, for example, has a prominent masthead and two text columns: a useful reminder that reading order matters. Publish the collection reference, tool version, processing date and reviewer alongside any released transcription, with unresolved readings visibly marked. An archive becomes more useful when automation increases access while the original evidence remains recoverable.
Sources and further reading
- National Archives of Australia, Digital Preservation Policy, version 1.4, 30 June 2020: content, metadata, integrity and preservation actions.
- Blnd Yaseen and Hossein Hassani, Making Old Kurdish Publications Processable by Augmenting Available Optical Character Recognition Engines, arXiv:2404.06101 (2024), especially pp. 20–22 and 26–27. Preprint.
- Sina Ahmadi, Daban Q. Jaff, Md Mahfuz Ibn Alam and Antonios Anastasopoulos, Language and Speech Technology for Central Kurdish Varieties, LREC-COLING 2024, pp. 10034–10045, especially 10036–10037 and 10040–10041.
- Geoffrey Haig, The Corpus of Contemporary Kurdish Newspaper Texts (2001), University of Bamberg repository record, DOI 10.48564/unibafd-v6rmw-cx940. Collection description consulted.
- Oral History Association, Oral History Best Practices: preparation, preservation, access and use.