Why is Arabic OCR harder than English OCR?

Four reasons compound. Arabic is cursive even in print, so letters join and character segmentation is not a solved preliminary step. Letterforms change shape by position, with initial, medial, final and isolated variants. Optional diacritics may or may not be present and may or may not be counted. And Arabic is right-to-left, so mixed Arabic and Latin content produces bidirectional sequences that frequently break reading order. The result is measurable: on the KITAB-Bench benchmark the median system records a 74.00% character error rate and only three of seventeen fall below 31%.

What is the most accurate Arabic OCR system?

On the KITAB-Bench image-to-text track, in a vendor-conducted evaluation by YaiGlobal, YaiGlobal Parser recorded the lowest character error rate of the seventeen systems compared, at 12.50%, followed by Gemini 2.0 Flash at 13.00% and AIN-7B at 20.00%. The evaluation has not been independently replicated, and it covers four of the benchmark's nine tasks. Accuracy rankings depend heavily on material type: systems that perform well on printed Arabic often fail on handwriting.

Can general-purpose AI models do Arabic OCR?

Some can, with caveats. Gemini 2.0 Flash is the strongest general-purpose baseline measured on KITAB-Bench image-to-text at a 13.00% character error rate, and GPT-4o records 31.00%. But general models degrade sharply on handwriting and on table structure, and they are prone to hallucination, producing fluent text that was never on the page. Several open-weight models exceed a 100% character error rate on this benchmark, which indicates generation rather than recognition.

How is Arabic OCR accuracy measured?

Primarily by character error rate and word error rate, both normalized edit distances where lower is better and neither bounded above by 100%. chrF, a character n-gram F-score, is a common auxiliary measure. Structural tasks use different measures: TEDS for table structure, mAP and F1 for layout detection, and MARS for end-to-end conversion. Whether a figure is macro-averaged or sample-weighted materially changes it.

Arabic OCR

Why the script is genuinely harder, how accuracy is actually measured, and what seventeen systems achieve when someone checks.

Arabic optical character recognition is frequently described as a solved problem by vendors and as an unsolved one by anyone who has tried to digitise an archive. The measurements settle it: on a peer-reviewed Arabic benchmark, the median system records a 74.00% character error rate and only three of seventeen fall below 31%. Arabic OCR is not a solved problem. It is a problem where a small number of systems work and most do not.

Why Arabic is harder than Latin-script OCR

Four difficulties compound, and none of them is a matter of needing more training data.

1. The script is cursive, even in print

Arabic letters join to their neighbours. There is no reliable whitespace between characters, so character segmentation is not a solved preliminary step the way it is for Latin print. Systems built on a segment-then-recognise architecture — which describes most traditional OCR engines — have a structural disadvantage before they see a single page.

2. Letterforms change shape by position

Most letters have initial, medial, final and isolated variants, and certain pairs form obligatory ligatures. The same letter can present four visually distinct shapes depending on where it sits in a word.

3. Diacritics are optional and inconsistently counted

Vowel marks (تشكيل) are omitted in most modern text and present in Qur'anic, poetic, pedagogical and some historical material. Whether an evaluation counts them as characters materially changes the resulting error rate — which means two benchmarks can report different numbers for the same system without either being wrong. See the glossary.

4. Right-to-left, and frequently bidirectional

Arabic runs right to left, but technical, legal and financial documents routinely embed Latin text, Western digits and Arabic-Indic numerals in the same line. Reading order is a separate failure mode from recognition: a system can transcribe every character correctly and still return them in an order that makes the document unusable.

What current systems actually achieve

These figures come from a vendor-conducted evaluation on KITAB-Bench, a peer-reviewed Arabic OCR benchmark (Heakl et al., Findings of ACL 2025) containing 8,809 samples across nine tasks. Baseline figures other than LlamaParse and Landing AI are quoted from the published benchmark paper.

Image-to-text character error rate, 3,760 samples. Lower is better.
System	Type	CER ↓
YaiGlobal Parser	Arabic document parser	12.50%
Gemini 2.0 Flash	General vision-language model	13.00%
AIN-7B	Arabic-specialised open model	20.00%
GPT-4o	General vision-language model	31.00%
Azure Document Intelligence OCR	Commercial OCR service	52.00%
Tesseract	Open-source OCR engine	54.00%
EasyOCR	Open-source OCR engine	58.00%
LlamaParse	Commercial document parser	69.00%
PaddleOCR	Open-source OCR engine	79.00%
Gemma3	Open-weight model	105.00%
Surya	Open-source OCR engine	495.00%

The distribution is the finding, not the ranking. Mean CER across the baselines is 126.19%; median 74.00%. Only three systems achieve a CER below 31%. The competitive group at the top is small, which should temper any reading of a leaderboard position as broad consensus — including ours.

Rates above 100% are not errors in the table. Character error rate is a normalized edit distance and is not bounded above; a system exceeding 100% is emitting more spurious content than there was text on the page. See why error rates exceed 100%.

Accuracy depends on material more than on system

A single headline accuracy number is close to useless for procurement, because the same system behaves very differently across material types. The clearest demonstration is handwriting:

Character error rate: full image-to-text task versus handwritten subsets.
System	Full task	Handwriting	Degradation
YaiGlobal Parser	12.50%	14.50%	+2.00 pts
Gemini 2.0 Flash	13.00%	22.00%	+9.00 pts
Azure Document Intelligence OCR	52.00%	76.50%	+24.50 pts
Tesseract	54.00%	75.75%	+21.75 pts
Qwen2.5-VL-7B	120.00%	160.75%	+40.75 pts

Systems tuned for printed Arabic do not transfer to handwriting. If your archive is handwritten, a benchmark score computed mostly on modern print tells you very little. More on Arabic handwriting recognition →

Recognition is not the whole job

Getting the characters right produces a flat stream of text. For most institutional use — search, retrieval, analysis, migration into a catalogue — that is not sufficient. Three further capabilities matter, and they are measured separately:

Structural tasks and their measures.
Capability	Measure	Why it matters
Table structure recognition	TEDS	A table with the right characters in the wrong cells is worse than useless for querying
Layout detection	mAP, F1	What makes coherent chunking possible for retrieval systems; without it a chunker severs headings from their sections
End-to-end conversion	MARS	The closest proxy for deployment: recognition, reading-order and structural errors all accumulate

The family ordering reverses between recognition and structure. LlamaParse records a 69.00% CER on raw transcription — ninth of seventeen — but 90.55 TEDS on table structure, second of fourteen. Purpose-built parsers carry an explicit geometric model of the page; general models must reconstruct it from pixels each time. A system's rank on one task predicts very little about its rank on another.

How to evaluate an Arabic OCR claim

Five questions that separate a measurement from a marketing number:

On what benchmark? A vendor-defined evaluation that the vendor wins is not evidence. A peer-reviewed third-party benchmark is a different object.
On how many samples, and which? Accuracy on 30 modern printed pages says nothing about a handwritten archive.
Against whom? Leading a field of three is not leading a field of seventeen.
Which metric, and macro or micro? The convention can move a headline figure by more than a point. Ours moves from 12.50% to 14.08%.
Who ran it, and has anyone replicated it? Ours was run by us. It has not been independently replicated, and until it is, that is a real limit on what it establishes.

One thing to check on any quoted figure. Character error rate is a ratio: 0.125 and 12.50% are the same number. A claim of 0.125% accuracy overstates by a factor of 100. It is a common enough slip that it is worth catching before it reaches a procurement document.

Send us your hardest page

Every number on this page is an average over someone else's documents. The only measurement that answers your question is one taken on your material. Send us the worst page you have — we will run it and send back the output next to the image.

Talk to YaiGlobal → See the full results

Benchmark results

Four tasks, 5,949 samples, 17 systems — including the three measures where a baseline leads.

Handwriting

Where the gap between systems is widest, and what a 14.50% error rate means in practice.

Glossary

CER, WER, chrF, TEDS, MARS, mAP — and when each one misleads.

Sources. Benchmark: Heakl, A., et al. (2025), KITAB-Bench: A Comprehensive Multi-Domain Benchmark for Arabic OCR and Document Understanding, Findings of ACL 2025, arXiv:2502.14949, DOI 10.18653/v1/2025.findings-acl.1135. YaiGlobal figures are from a vendor-conducted evaluation on four of the nine tasks, 5,949 samples, not independently replicated — full methodology and limitations.