OCR for Non-English Text: Hindi, Chinese, Arabic and More

OCR guide · 5 min read

On-device OCR can pull text out of a photo, a screenshot, or a scanned page in seconds — but only if it knows which script it is looking at. A model trained to read Latin letters will stare at a line of Hindi or Chinese and confidently return nonsense. The single most important setting for accurate results is the recognition language, and it matters far more for non-English text than most people expect.

This guide explains why script choice matters, how different writing systems behave, and what to check when the extracted text comes out wrong.

Why the recognition language matters so much

OCR is pattern matching against a trained alphabet. When you tell the engine "this is English," it compares each shape against the 26 Latin letters, digits, and common punctuation. Point that same English model at Devanagari or Han characters and it has no matching shapes to fall back on — so it guesses, mapping curves and strokes onto whatever Latin letters look vaguely similar. The output looks like random symbols, and no amount of re-scanning fixes it, because the problem is the alphabet, not the image.

Choosing the correct language loads a model trained on the right script, the right character set, and often the right typical layout. That one choice usually makes the difference between clean text and garbage. In Textquill you pick the language before extracting; English is bundled and works offline immediately, while the other 15 languages download their data once and then run offline too.

How script differences affect OCR

Writing systems are not just different letters — they differ in structure, and each difference is something the OCR model has to be trained for.

• Devanagari (Hindi, Nepali): Letters hang from a top horizontal line (the shirorekha), and vowel signs attach above, below, or beside a consonant. The engine has to segment these stacked marks correctly, so a Devanagari-trained model is essential.
• Han characters (Chinese, Japanese): Thousands of distinct characters, often with no spaces between words. The model recognizes whole characters rather than assembling letters, and Japanese mixes Kanji with Hiragana and Katakana in the same line.
• Arabic: Written right-to-left, and letters change shape depending on their position in a word (isolated, initial, medial, final). Text direction and letter-joining both have to be handled or the words come out scrambled.
• Diacritics (Vietnamese, Turkish, and accented European languages): Small marks above or below letters change the meaning entirely. Vietnamese in particular stacks tone marks on vowels, so the model must preserve them, not smooth them away.
• Thai: No spaces between words, and vowels and tone marks sit above and below the consonant line, which makes line segmentation harder than in spaced scripts.

How automatic detection helps — and where it stops

Automatic language detection is genuinely useful when you don't know what you're looking at, or when you process many images from different sources. It samples the shapes in the image, guesses the most likely script, and applies the matching model. For clean, single-language images it usually gets it right.

It has real limits, though. Detection works best with a decent amount of clear text; a few words, a stylized font, or a low-resolution crop gives it little to go on. Scripts that share characters — Chinese and Japanese Kanji, or the several languages that use Latin letters — are easy to confuse. And auto-detection generally commits to one language for the whole image, so a bilingual sign can trip it up. When you already know the language, selecting it manually is more reliable than letting detection decide. Textquill offers automatic detection for convenience, but treat the manual language list as your accuracy tool.

Tips for accuracy by script family

1. Set the language first, then extract. This is the biggest lever. If you know the text is Korean, choose Korean rather than relying on detection.
2. Give dense scripts more resolution. Han, Devanagari, and Thai pack a lot of detail into each character. A larger, sharper image helps far more than it does for spaced Latin text. Zoom in or use a higher-resolution capture.
3. Keep the text upright and reasonably straight. Skewed or rotated lines hurt every script, and especially the connected letterforms of Arabic.
4. Prefer good contrast over color. Dark text on a plain light background (or the reverse) reads more reliably than text over a busy photo.
5. Crop to just the text you need. Area select (Alt+Shift+S in Textquill) lets you isolate one block, which removes distracting shapes and lets the model focus on a single language.

Handling mixed-language images

Real documents are messy: an English caption under a Chinese heading, a Hindi paragraph with English brand names, a menu in two languages. Because most OCR runs assume one primary script, mixed images are where results wobble. The practical fix is to process one language at a time. Use area select to grab the Hindi block with Hindi chosen, then grab the English block with English chosen. Two clean passes beat one confused pass. If the languages are interleaved word by word, pick the script that carries the most important content and accept that stray words in the other script may need a quick manual touch-up.

When extraction comes out wrong: what to check

If the output looks like random characters or empty gaps, work down this list before blaming the image:

• Wrong language selected. This is the number one cause by a wide margin. Reading Arabic with English (or Chinese with Japanese) selected produces garbage. Switch to the correct language and extract again.
• Auto-detection guessed wrong. If you used automatic mode on a short or stylized snippet, override it with the specific language.
• Language data not downloaded yet. For non-bundled languages, make sure the data finished downloading the first time; after that it works offline.
• Image too small or blurry. Re-capture at higher resolution, especially for dense scripts.
• Mixed languages in one grab. Split the image and run each language separately.

Ninety percent of "OCR doesn't work for my language" reports come down to that first item. Match the recognition language to the script on the page, and most of the other problems disappear.

FAQ