By Meside AI · Last updated: 2026-08-06
The most common reaction to a finished English-to-Chinese book translation is not "is this any good?" It is "where did the rest of my book go?"
The file is smaller. The reader's progress bar moves faster. A chapter that ran eleven screens now runs four. Something looks lost.
Nothing is lost. Chinese is simply a much denser writing system than English, and the size of that difference is larger than most people expect. We translate EPUBs for a living, so instead of estimating it, we measured it across every book our pipeline has translated into Chinese.
Every EPUB in our system with a completed Chinese translation: 17 completed translation runs covering 11 distinct books, 44,543 translated text fragments, 6,769,384 characters of English source.
The gap between 17 and 11 is worth stating rather than rounding away: our smallest book is a short file we re-run as a regression test, so it accounts for seven of the seventeen runs. It contributes under 2% of the characters measured, but "17 books" would have been the wrong number to write down.
All measurements are on text with the HTML markup stripped out. That matters more than it sounds — an EPUB is a ZIP archive of XHTML files, and the tags are identical on both sides of a translation. Leave them in and they pad both columns, dragging every ratio toward 1 and hiding the effect entirely.
The books are a mix of technical references, business and startup non-fiction, and general non-fiction, ranging from 16,595 to 1,587,070 characters. We are not naming them — they are user-uploaded files. Where we need something you can check yourself, we use a public-domain book, noted below.
Across all 44,543 fragments:
| Measure | Value |
|---|---|
| English source characters | 6,769,384 |
| Chinese output characters | 2,351,707 |
| Chinese ÷ English | 0.35 |
A 500,000-character English book becomes roughly 175,000 characters of Chinese. That is not compression, summarising, or truncation — it is what the same meaning costs in a logographic script:
school 6 characters → 学校 2 characters
translation 11 characters → 翻译 2 characters
the weather 11 characters → 天气 2 characters
Multiply that across every noun in a book and a third is exactly where you land.
The number is also remarkably stable. Per book, across a 96× range in size and four different genres:
| Book size (source chars) | Fragments | Chinese ÷ English |
|---|---|---|
| 1,587,070 | 11,286 | 0.335 |
| 1,159,920 | 6,456 | 0.351 |
| 834,520 | 6,474 | 0.356 |
| 606,551 | 4,221 | 0.361 |
| 514,485 | 1,378 | 0.316 |
| 203,627 | 2,525 | 0.329 |
| 16,595 | 255 | 0.351 |
Every book in the set lands between 0.316 and 0.380 — including the 16,595-character regression file, which is reassuring in its own right: seven separate runs of the same text landed within 0.347 to 0.355 of each other, so run-to-run variation is far smaller than book-to-book variation.
Going fragment by fragment through the 12,678 substantial prose paragraphs (over 200 source characters), the median is 0.308, with half of all paragraphs falling between 0.282 and 0.344.
So this is predictable enough to plan around. If your Chinese translation came back at roughly a third the character count, it is behaving exactly like every other book we have measured.
The one place the ratio moves is fragment length, and it moves in a direction worth knowing about:
| Source fragment length | Fragments | Chinese ÷ English |
|---|---|---|
| Under 20 characters | 9,330 | 0.81 |
| 20–79 characters | 10,468 | 0.42 |
| 80–199 characters | 6,978 | 0.38 |
| 200–599 characters | 10,923 | 0.33 |
| 600+ characters | 1,794 | 0.35 |
Very short fragments — chapter numbers, table cells, figure captions, index entries, a single bolded term — hardly compress. Some shrink as usual:
Chapter 12 10 characters → 第12章 4 characters
But plenty of short fragments are proper nouns, version numbers, code identifiers or measurements, and those carry straight across unchanged at a ratio of 1.0.
This is why the shrink is so visible in a reader: it is concentrated in exactly the prose you actually read. Your table of contents looks normal. Your chapters look half gone.
Same pipeline, same measurement, other target languages:
| Target language | Distinct books | Fragments measured | Output ÷ English source |
|---|---|---|---|
| Chinese (Simplified) | 11 | 44,543 | 0.35 |
| French | 1 | 510 | 1.19 |
| Spanish | 1 | 9,149 | 1.03 |
| Turkish | 1 | 1,399 | 0.96 |
Be careful with the bottom three rows. Only the Chinese row has a real spread of books behind it; the others are a single title each, measured across repeated runs, which is why their fragment counts look larger than one book's worth. Treat them as direction, not as measurement.
The direction is unambiguous, though, and it is the part worth taking away: European targets land near or above parity, and Chinese is a different regime entirely. If you have only ever translated a book into a European language, your intuition about output length does not transfer.
The Spanish row is reproducible. It is Project Gutenberg #43936, The Wonderful Wizard of Oz, which is public domain: 233,240 characters of English across 1,948 fragments, coming back as roughly 241,000 characters of Spanish. Download it and count for yourself.
Your book is not truncated. Check completeness by structure, not by length: does the translated EPUB have the same chapters, the same table of contents, the same number of images? A missing chapter is a real failure. A chapter that is a third the length is arithmetic.
Your reader's page count and time estimates will change. Most e-readers estimate reading time from character or word counts calibrated on Western scripts, so they will badly underestimate how long the Chinese edition takes to read. This is a property of the reader, not the translation.
What you pay does not shrink. The price is computed from the source file you upload — $0.005 per 1,000 characters, $1.99 minimum per book — not from what comes back. That is deliberate: it means you know the number before any work happens, and it does not change based on which language you pick. See pricing for the full breakdown.
Consider keeping the original alongside. Because the Chinese is so much shorter, a bilingual EPUB — original paragraph followed by translation — costs far less added length than it would in French or Spanish. If you are reading to learn the language, this is the format worth asking for.
We treat Simplified Chinese and Traditional Chinese as two different target languages, not as one with a font switch.
They are not the same text with different glyphs. The vocabulary itself diverges, so a character-by-character conversion gets the words wrong, not just the shapes:
software mainland 软件 Taiwan / HK 軟體
network mainland 网络 Taiwan / HK 網路
print mainland 打印 Taiwan / HK 列印
Run a naive converter and you get text that is technically readable and obviously machine-made to anyone from Taiwan or Hong Kong. Pick the one your readers actually use — all supported target languages are listed here, 16 in total.
If you are considering doing this yourself with an API key and a script, translation quality is not what will stop you. Modern models handle English-to-Chinese prose well.
What will stop you is that a book is not a document. It is thousands of separate fragments spread across dozens of XHTML files inside a ZIP, each of which has to come back with its markup intact and get put back exactly where it was. The books above split into anywhere from 255 to 11,286 fragments, and the fragment count does not track length — one book in the set is 2.5× longer than another and splits into half as many pieces.
We wrote that up with the full numbers in how many tokens is a book, and the engineering side in why building an EPUB translator is harder than you think. The short version: budget your effort for structure and scale, not for the translation step.
There are honest alternatives. If what you want is to read Chinese-language web pages and the occasional EPUB in a browser, a reader-side extension such as Immersive Translate is a sensible place to start and supports bringing your own model. If you have one short book and an afternoon, a script plus an API key genuinely works.
What we built EPUB Translator for is the other case: a whole book, in one pass, that still opens as a valid EPUB at the end — segmentation, per-fragment retry, structural repair and validation. If that is the job, start from the pricing page and you will see the exact cost for your file before anything runs.