Arabic OCR is solvable, but most tools either silently fall back to Latin (turning every page into garbled placeholders) or get the characters right but break the layout. This guide explains why Arabic is harder than English to OCR, what to check on the output, and how to handle the common variants — Modern Standard Arabic, Persian/Farsi, and Urdu.
Why Arabic OCR Is Different
Right-to-left reading order
Arabic flows right to left, but bidirectional text mixes Arabic with embedded Latin (English brand names, numerals written left-to-right). Recognizers have to identify the script of each run and emit the right reading order — get this wrong and a phone number ends up reversed.
Contextual letter shaping
Most Arabic letters take four shapes depending on position: isolated, initial, medial, and final. The letter ع (ayn) looks like a curl on its own, an upward hook at the start of a word, a flattened curl in the middle, and a deep tail at the end — all the same letter, all visually distinct. An OCR engine that hasn't been trained specifically on Arabic will treat these as four different characters and the text won't copy back as a real word.
Diacritics (harakāt)
Religious texts and educational material include vowel marks (fatha, kasra, damma, sukun, shadda) sitting above and below the consonant line. These are tiny — easy to miss at low scan resolutions — but they change meaning. Strip them and you change the word.
Connected script with kashida
Letters in a word join, and printers stretch the connecting line (kashida) for justification. The recognizer has to ignore the stretching and treat the kashida as part of the same letter rather than a separate stroke.
Persian, Urdu, and Other Arabic-Script Languages
Persian (Farsi), Urdu, Pashto, Sindhi, and Uyghur all use Arabic-derived scripts but add letters Arabic doesn't have — پ (pe), چ (che), ژ (zhe), گ (gaf) for Persian; the additional retroflex letters for Urdu. A recognizer trained only on Modern Standard Arabic will misread these as the closest Arabic letter and you'll get the wrong word.
Urdu in particular is usually written in Nastaliq calligraphy — a slanting, cascaded style that's much harder to recognize than the upright Naskh used for Arabic. Some engines support Urdu only in Naskh; if your document is Nastaliq, accuracy drops substantially.
What to Check on the Output
- Spot-check on the page. Open the searchable PDF, select a known word with the cursor, copy, paste into a text editor. If the pasted text reads correctly in Arabic, the OCR worked. If you see runs of
?or Latin placeholders, the embedded font lacks proper Arabic mappings. (This is a common problem — many libraries default to a Latin-only font for the OCR text layer.) - Check directionality. When you copy a phone number embedded in Arabic text, does it stay left-to-right? Numerals should preserve their LTR order even inside an RTL paragraph.
- Look for diacritics. If the original has harakāt (vowel marks) and the output doesn't, the recognizer dropped them. For most legal/news documents this is fine; for religious or educational text it isn't.
- Search for a long word. Search a 5+ letter word via Ctrl-F. If it finds the word, the contextual shaping was handled correctly — letters were stitched back into Unicode codepoints rather than left as glyph IDs.
Common Document Types That Need This
- Legal records — Arabic court judgments, land deeds, contracts. These often have formatted columns and stamps, so layout-preserving extraction matters.
- Religious texts — Quran printings, Hadith collections, classical commentary. Diacritics matter; pick a tool that preserves them.
- Government forms and certificates — usually a mix of Arabic body and Latin numerals/codes. Bidirectional handling is the failure mode.
- Newspaper archives — multi-column, dense, often decades old. Page resolution and skew handling become the bottleneck.
convert2searchablepdf for Arabic
We use Google Cloud Vision's document text detection, which handles Arabic, Persian, and Urdu (Naskh) reliably. The text layer is drawn with a Unicode-aware font (Noto family) that includes the full Arabic block plus contextual shaping forms — so when you copy a paragraph out, it pastes as real Arabic, not placeholder characters. The layout-preserving text extraction (the "Show Text" button after conversion) keeps right-to-left order intact.
Try it with your own document — drop the PDF on the homepage and you'll get both the searchable PDF and a plain-text export back in seconds.