Kindi Labs | Teaching machines to read Arabic

01Kindi OCRv1.0

Printed Arabic

Kindi OCR reads scanned Arabic newspapers and books, and returns text you can search and correct.

1
Block by block
The page is read one text block at a time, so columns and headlines don't run together.
2
Line by line
Each block is read line by line, and each word joined up the way it's written.
3
Then we measure
We compare the output with the true text, character by character, on 90 newspaper blocks.

Extracted textCharacter error rate0.89%

الحركة الدائمة

فقلنا وهل تدّعي ان الآلة التي تشير اليها ترفع الماء من نفسها من غير قوة تساعدها على رفعه.

فقال كلاً بل اننا نضع ماءً في جانب منها فتدور به وتغرف الماء من النيل وترفعه. فقلنا ومن اين

تأتي بالماء لا ترفعه انت من النيل لتضعه في الآلة فارو الارض به بدلاً من وضعه في الآلة ثم

انك اذا فعلت ذلك لم تخسر شيئاً من الماء الذي رفعته بيدك واما اذا وضعته في الآلة فيستحيل

ان يرفع بها ما يساويه تماماً في كميته او في المسافة التي ارتفع اليها لان الآلة تخسر شيئاً من قوة

For illustration: a page of Al-Muqtataf, June 1899 (public domain, via the Internet Archive). This page is not part of our test; the percentage is the result on the test described below it. Internet Archive

On 90 newspaper blocks, each scored as a whole blockLower is better

On 90 newspaper blocks, each scored as a whole block. MDP-90. The whiskers at each bar show the 95% interval.
System	Character error rate	95% interval
Kindi OCR	0.89%	0.6–1.2
Mistral OCR v4.1	1.29%	1–1.6
deep.sa OCR v2.0	2.63%	1.9–3.7
PaddleOCR-VL 1.6	5.4%	4.2–6.5
Tesseract	6.53%	4.5–9.3
	0%2%4%6%8%10%

MDP-90. The whiskers at each bar show the 95% interval.

Everything about Kindi OCR

02Kindi HTRUnder development

Handwriting

Handwriting is harder than print: every writer has their own hand, and every script has its own rules. Kindi HTR reads five historical Arabic scripts, and we score it on the same lines every time.

Where each interval sitsOur upper bound, 9.26%, sits under the closest system's lower bound, 20.89%. The intervals don't overlap.

Script by script: Kindi HTR against HATFormer

Script by script: Kindi HTR against HATFormer
Item	HATFormer	Kindi v0.5 preview
Oriental	32.0%	3.7%
Ruq‘ah	17.9%	5.8%
Nastaliq	20.7%	8.2%
Ta‘liq	13.4%	8.8%
Naskh	21.2%	9.8%
Maghrebi	33.9%	10.4%
Modern	41.4%	15.3%

Same lines, same score, for every script, on a wider test of 8,136 lines. Values are character error rates. On modern handwriting, three other systems we tested read better than Kindi HTR.

Everything about Kindi HTR

03laya-ara-rag

Arabic search

Before an assistant answers, it has to find the right passage. Give laya-ara-rag a question and up to 12 passages, and it picks the one that answers.

We trained it on MIRACL-ar alone. It had never seen the other four sets.

The question

أيّ خطٍّ شاع في الكتابة اليومية لسرعته وبساطته؟

النسخخط واضح متناسق، شاع في نسخ الكتب والمصاحف، ومنه أُخذت حروف الطباعة العربية.
الثلثخط فخم كثير التشكيل، يُكتب به في عناوين الكتب وعلى جدران المساجد.
الكوفيخط هندسي حادّ الزوايا، كُتبت به المصاحف الأولى.
الرقعةخط سريع قصير الحروف بلا زخرفة، شاع في الكتابة اليومية والمراسلات.First pick
الديوانيخط متداخل الحروف، كُتبت به المراسيم في دواوين الدولة العثمانية.
النستعليقخط مائل رشيق، ارتبط بالأدب الفارسي ودواوين الشعر.
المغربيخط أهل المغرب والأندلس، تُنقط فيه الفاء من تحت.
التعليقخط فارسي سبق النستعليق، استُخدم في المراسلات الرسمية.

Illustration: we wrote the question and passages, and the scores are for display only.

First-pick accuracy, before and after our training

First-pick accuracy, before and after our training
Item	Base model	laya-ara-rag
MIRACL-ar (trained on)	15.3%	58.8%
Mr.TyDi-ar	17.6%	63.2%
SadeemQuestion	16.8%	87.1%
MLQA-ar	13.3%	58.2%
XPQA-ar	21.3%	66.5%

The base is laya-multilingual before our training. Share of questions where the first pick was the right passage.

Open on Hugging Face, Apache-2.0

Teaching machines to read Arabic.

Printed Arabic

Block by block

Line by line

Then we measure

Handwriting

Arabic search

Run the models on your own pages.