Blog

Screenshot OCR in 11 Languages: Which Work, Which Don't

Written by InSnap · Published October 30, 2026 · ocr, languages, japanese, korean, russian, screenshot

We tested offline OCR on 11 interface languages. English, Chinese and Latin-script languages read almost perfectly; Japanese, Korean and Russian failed. The numbers.

InSnap's offline OCR read eight of the eleven interface languages we tested almost perfectly, and failed on three: Japanese, Korean and Russian. Apple's text-recognition API, tested the same way, read all eleven at 98% or better.

So the answer to "does screenshot OCR work for my language" is not a yes or no. It depends on the script, and the gap is large enough that it is worth checking before you rely on a copy.

How we tested

For each language we took six real sentences from InSnap's own interface, in that language's translation, and rendered them in the system UI font at normal size. That gives a clean test sheet with exactly known text. We made each sheet at Retina (2x) and at 1x.

Two engines read every sheet:

• InSnap's OCR, which runs fully on your computer with a bundled model and is identical on Mac and Windows.
• Apple's Vision text recognition, the recognition API built into macOS, in accurate mode with automatic language detection. We tested the API directly, not a particular app that uses it.

The score is the average similarity between what the engine read and the real sentence, after ignoring spaces and case. 100% means every sentence came back exactly.

The results

Retina and 1x sheets scored within a point of each other, so resolution was not the issue.

What a failure looks like

The three weak languages are not read badly. They are read as something else. Here is one of the test sentences, which a reader sees as perfectly clear text:

Japanese keeps its kanji but swaps many kana for look-alike kanji. Korean comes back as a mix of Chinese characters, digits and punctuation. Russian is read as Latin letters and digits that resemble Cyrillic ones. In all three cases the output looks like text, so nothing warns you. (The Korean output is shortened here.) Traditional Chinese is mostly fine, but a few characters came back in their Simplified forms, so proofread names and quotes.

What this means for you

• English, Chinese and Latin-script languages: InSnap is accurate on clean text. Check numbers and proper names anyway; see how to get better OCR results.
• Japanese, Korean and Russian: do not copy from InSnap's OCR without checking every line. In our test Apple's recognition API handled all three, so on a Mac a tool that uses it is the better choice for those scripts. We did not test Windows' built-in recognition.
• Mixed text: select one script at a time where you can.

InSnap uses one bundled model on both Mac and Windows, so the result is the same on either, and nothing is uploaded. The trade-off is that its language coverage is fixed by that model.

What the test does not show

• Clean text only. Real screenshots carry compression, small sizes and busy backgrounds, and every engine scores lower there.
• Six sentences per language. Enough to show gaps of 60 to 90 points, not enough to rank two engines that both score above 97%.
• One font and size. Stylized fonts and handwriting are harder for any engine.

FAQ

Which languages can InSnap's OCR read?

In our test it read English, German, French, Spanish, Portuguese, Italian and Chinese (Simplified and Traditional) at 97% or better. Japanese, Korean and Russian were not readable.

Why does Russian text come out as Latin letters?

The recognition model was not built to read Cyrillic, so it picks the closest shapes it knows. Many Cyrillic letters look like Latin letters or digits, which is why the output looks like garbled English.

Does a larger screenshot or higher resolution help?

Not here. Retina and 1x test sheets scored within a point of each other. The limit is the model's language coverage, not the image quality.

Is the operating system's built-in OCR more accurate?

For Japanese, Korean and Russian, yes: Apple's Vision API read all three at 99% or better in our test. For English, Chinese and the Latin-script languages the two were equal within a few points.

More from the blog