How to Translate a Scanned Japanese PDF to English

To translate a scanned Japanese PDF to English, you need three things working together: OCR that reads Japanese scripts and vertical text correctly, translation that sees whole passages rather than isolated lines, and a way to put the English back onto the page so tables and stamps stay where they were. Generic translation tools reliably fail at the first step — not because Japanese OCR is impossible, but because Japanese pages violate the assumptions those tools were built on. Once you understand which assumptions break, the fix is straightforward.

Why Japanese scans are a special case

Vertical text read in the wrong order

A large share of Japanese documents — books, official notices, older contracts, most anything traditional — are set in tategumi: characters run top to bottom within a column, and columns proceed right to left across the page. Generic OCR assumes horizontal lines read left to right and top to bottom. Point it at a vertical page and it does one of two things: reads across the columns, producing a shredded interleave of unrelated sentences, or detects each column as a stack of one-character "lines." Either way, the text that reaches the translator is scrambled before translation even begins, and no translation engine can recover meaning from characters in the wrong order.

Three scripts on one line

Japanese mixes kanji, hiragana, and katakana in a single sentence, often with Latin letters and Arabic numerals thrown in for names, model numbers, and dates. OCR trained mostly on one script confuses visually similar characters across scripts — a katakana ニ misread as the kanji 二, カ as 力 — and each substitution is a small landmine, because a one-character OCR error in Japanese can flip the meaning of a word entirely rather than just misspelling it.

Furigana doubles the text

Formal and educational documents print furigana — small kana reading guides — alongside kanji, above them in horizontal text or to the right of the column in vertical text. To OCR that doesn't know what furigana is, these are just more characters, and they get interleaved into the main text: every annotated word effectively appears twice, once as kanji and once as its phonetic reading, welded together mid-sentence. Translators then produce doubled or garbled English from the doubled Japanese.

Hanko seals and warichu

Japanese official documents are validated with red hanko seals stamped directly over text — often over the exact name or date you most need to read. OCR either fails on the overlapped characters or hallucinates replacements. Contracts and legal texts add warichu: parenthetical notes set as two half-height lines squeezed inside one line of the main text. Generic OCR reads the two miniature lines as if they were part of the surrounding sentence, splicing the annotation into the clause it annotates.

Why line-by-line copy-paste is brutal for Japanese

The standard folk workflow — OCR the page, copy the text into a translator, paste the output into a document — is painful in any language and close to unworkable in Japanese. Japanese has no spaces between words, so when OCR emits hard line breaks (and for vertical text, a break per column or worse), the translator can't tell whether a break falls inside a word. Sentences also routinely continue across columns, so translating line by line hands the engine fragments with no grammatical subject — which Japanese omits freely when context supplies it. The result is a translation where every fragment is individually plausible and the passage as a whole is wrong. To do it properly by hand you'd have to rejoin every column into full sentences first, for every page.

A workflow that actually works

1. 1Start from the best scan you can get. 300 DPI or better, pages straight, even lighting if you scan with your phone. Japanese OCR is more sensitive to quality than Latin-script OCR because so many characters differ by a single stroke. Clear printed pages work well; heavy handwriting — common on older Japanese forms — remains unreliable for any OCR, so flag handwritten sections for human review rather than trusting the output.
2. 2Use OCR that understands Japanese layout. The tool must detect vertical writing, read columns right to left, keep furigana separate from base text, and cope with seals overlapping characters. This single step is where generic tools lose; everything downstream depends on the text arriving in the right order.
3. 3Translate passages, not lines. Modern AI translation handles Japanese-to-English well when it sees complete paragraphs with context — it can resolve dropped subjects and honorific register. It handles isolated column fragments badly. The OCR step must therefore reassemble full sentences before translation.
4. 4Re-typeset the English into the layout. Japanese is compact; its English translation typically runs longer, and vertical source text usually needs to become horizontal English. The final step places the translation back where the original text sat — resized and reflowed to fit — so tables, seals, figures, and numbering stay put.

Doing it in one pass

Reglyph was built around exactly this pipeline for scanned PDFs, phone scans included: OCR reads the page, the original text is erased from the page image, and the English translation is typeset back in place, so the output is a PDF that still looks like the source document. That holds for horizontal pages; a vertical (tategumi) page can be read, but its English comes back as horizontal text and will not match the original layout. Hanko seals, ruled tables, and figures stay where they were because the page image underneath is preserved. For documents where you need to verify the translation against the original — common with Japanese, where a reviewer may want to check a specific clause — the bilingual side-by-side export puts source and translation next to each other page by page. It works in the browser, including on mobile, you get 5 free pages a day with no credit card, and paid use starts at $3.

What to check in the output

However you translate, audit a Japanese result differently than a European-language one:

• Reading order — pick one vertical page and confirm the English follows the columns right to left. Scrambled order means the OCR step failed, and no downstream fix helps.
• Names and dates under seals — hanko often sits on the signature block. Verify those characters against the source rather than trusting OCR through red ink.
• Numbers — Japanese documents mix Arabic numerals, full-width numerals, and kanji numerals (三十 for 30), sometimes in the same table. Confirm amounts survived.
• Era dates — 令和6年 is Reiwa 6, i.e. 2024. Check whether era years were converted or left literal, and be consistent about which you want.
• Doubled words — a phrase that appears twice in slightly different form is usually furigana bleeding into the main text upstream.

A Japanese scan that comes back in the right order, with seals still in place and numbers intact, is a solved problem. The whole game is choosing tools that treat Japanese layout as a first-class case instead of an afterthought.