How to translate a book without losing voice
Readers can’t tell an AI translation from a human one, but they can tell when the voice goes flat. Where AI book translation loses the author’s voice, what a high-quality translation balances, and the five steps we use to keep the voice: a three-page sample, a style guide, a glossary with translation memory, scene-sized calls, and AI-assisted review.


On this page
In a 2026 study of fifteen translated novels by Ferstler et al., readers who read 12–120 books a year tried to tell which excerpts came from a human translator and which were translated by AI. They guessed right 34 times out of 60: basically, a coin flip.
Experienced readers could not detect AI. But they could detect bad translation.
When they compared 772 passages side by side and picked what they actually liked, without knowing which was which, readers preferred the human translation 522 times, in 67.6% of cases. Readers felt when the voice and natural rhythm were off. And where human translators lost to AI, it was because they broke the text flow, which made the translation hard to read.
We initially built Transept to translate our own fiction writing, as authors. We felt this AI translation effect right away: an author sees at once how puns and wordplay get lost.
And as developers, we worked hard to figure out how AI translation loses your voice, and how to recover it and bring it to a native level using your unique author knowledge.
To translate a book with AI without losing voice, dial in your style on a 3-page sample first, then enforce it with a style guide, glossary, and translation memory. Translate in scene-sized chunks instead of paragraph by paragraph, and run AI-assisted review so you only fix what actually broke.
Below, I walk through all five steps in Transept, though you can recreate the same flow with your own tools.
What makes AI book translation lose original voice?
AI book translation loses original voice in four ways: literal idioms, copied source syntax, drifting character voices, and wordplay flattened into safe phrasing. Here is what academic research and our own Transept experiments show:
Idioms translated word for word
Models translate the dictionary surface of an expression instead of how a person actually speaks. In Wiesław Myśliwski’s Needle’s Eye, a common Polish phrase for failing legs came out of the AI pipeline as “his legs refuse obedience and his eyes no longer come together in seeing”, while human translator Bill Johnston wrote “when your legs won’t do what they’re told to and your eyes can’t make out the world.”
Sentence structure copied from the original language
LLMs copy the clause order and sentence length of the source instead of rebuilding it for the target language. Readers in the same study found the AI version of Needle’s Eye “very choppy”, with “bizarre punctuation choices”, and said it was hard to tell who was speaking. A human translator splits and reorders the clauses so the English actually breathes.
Character voice drifting from chapter to chapter
Without book-level memory, characters’ unique voices slip between scenes or vanish into a safer register. When we localized our own translation assistant into Ukrainian, her signature smug mood disappeared completely. The literal Ukrainian word («самовдоволена») reads as a pure insult rather than fond self-satisfaction. So the model, trained to be harmless and helpful, quietly made her overall demeanor kinder and more polite than in English.
Wordplay and sharp style flattened into safe phrasing
Models pull puns, coined terms, and deliberate human choices back toward high-probability phrasing. In Guerberof-Arenas & Toral’s study of Kurt Vonnegut’s “2BR02B”, every reviewer rated the translations made from scratch as the most creative and the raw machine output as the least: the machine “renders literal solutions to translation problems”. Given “What man in my shoes wouldn’t be happy?”, it kept the shoes in Catalan («Quin home de les meves sabates no estaria content?»), and all three reviewers marked it a major error.
What makes a book translation high-quality?
Think about the most obvious examples, like character names, or worldbuilding elements like toponyms and magic terms. Not only did they not exist before as literal strings, but they carry a lot of information on different levels: literary references, style register, associations, and cultural connotation. For example, «Винагорода» as a sweets brand would read as a down-market Soviet artifact in Ukrainian, while “Bounty” (the exact English match) would not.
A good translator has to either reinvent the term in the target language or leave it untranslated to preserve the register. A notorious example is the panned russian translation of Harry Potter, where translating character names ruined the British vibe of the magic. Whereas the Ukrainian translation by Viktor Morozov kept most proper nouns intact, localized the stylistic texture instead, and became one of the most praised.
Translating successfully, therefore, requires AI to differentiate between inventions and idioms, and either re-create them in the target language or pick the equivalent. Each time, it requires a choice, sometimes changing large spans of text.
That is what we made Transept for: so the author chooses once, and the AI carries that creativity across the rest of the book.
How do you translate books with AI to get native rhythm and style?
Getting native rhythm and style takes five steps: nail three pages first, write a style guide, build a glossary and translation memory, translate by scene, and review with AI. I started building this workflow for my own books in 2023 and still refine it with every new translation.
This guide assumes that you know the target language, at least well enough to make a confident judgment as a reader.
If you don’t, you can still follow the same flow: generate several options for your 3-page sample, have a strong AI pick the winners to keep the voice consistent, and bring in a native reader via review links before you publish.
1. Iterate on translation style until you find the winner
Before you translate the whole book and spend money on it, you need to develop a consistent style that actually feels like you. It’s easier for AI to leech off a good reference than to invent one on the fly.
Start with a short slice of your text, up to 3 pages or 900 words. Your “golden slice” should contain:
- A dialogue piece, preferably with conflict, where characters show up
- A dynamic scene snippet
- An expository, descriptive scene
- What you think makes your writing yours: some poetry, some puns and wordplay, or just a snippet that you hold dear.
This will give a sufficient range for you to test everything.
Then, you need to generate the translations with AI. What’s more important, you need to give AI feedback, and iterate until your AI gets it.
I created (and fought to keep) several tools in Transept to speed up this process.
-
Sentence alternatives suggest options for a selected sentence, DeepL-style. They can also suggest merging or splitting sentences if it improves the flow.
-
Regenerate creates a new version of the paragraph, considering your comments.
-
Speaking of comments, you can annotate the page, and any tool, like proofreading, regeneration or glossary building, will be able to see them.
-
I also like speaking to Literess, dumping my stream-of-consciousness critique on her until she gets it right. It helps to get breakthroughs when I am stuck.
This is the single most important moment of book translation with AI. Don’t rush it. Here, you keep your unique human contribution that makes it stand out, and lay the foundation to speed up the following steps.
Read nextHow human and AI translation differ, and what makes readers like both? A cross-corpus empirical re-analysis and LLM-as-a-Judge meta-evaluationAI keeps the information; humans spend it on voice. We measured where the two differ across four public corpora, and what readers and experts reward in each.
2. Generalize the best style into your own style guide
Now that you have created a sample of your style, you need to generalize it into guidance for AI to preserve your voice.
You should include the style guide during each generation, and also when editing or discussing your book. Preferably, if your tool allows, it should not go into the system prompt to leave AI some room for interpretation.
The style guide should include:
- The basic settings: tone, register, who your reader is, how characters address each other (a formal or informal “you”), sentence rhythm, how far to localize cultural references, profanity, foreign words, quotation marks, and number and date formats.
- Do and don’t rules about patterns across the whole text: “Keep the narrator’s sentences short”, “Don’t modernize period slang”. Never about single words or names: those belong in your glossary.
- Notes on the voice: a few paragraphs of prose on the narrator’s voice, how dialogue differs from narration, the era’s conventions, humor and irony, the texture of the sentences. Whatever an editor would tell a new translator on their first day. This is the most valuable part.
- Reference samples: short passages from your winning style sample, quoted word for word next to the translation you approved. These can be a typical dialogue exchange or a paragraph of narration with the right rhythm. Samples teach the voice better than any rule.
Characters’ voices go into the notes and the samples: a sentence on how someone speaks, and a dialogue exchange that shows it.
In Transept, you can build a custom style guide for each target language (or start from a preset if you are new to it). It can read your existing translations to pull out your stylistic patterns, and update the style guide as you go, which matters if you write and translate chapter by chapter. You can also workshop it with Literess when you want a second pair of eyes to check whether a phrasing is idiomatic or invented.
3. Use glossary and translation memory for your book
Professional translation, even before AI, has been based on two pillars: glossaries and translation memory (TM). To make your book sound consistent and native, you need to have both.
Creating a glossary is my favorite part of starting any translation. There is so much lore and artistic purpose behind names, terms, and toponyms, and laying them out forces me to think through the story more clearly. I have often returned to the original and improved it after building the glossary for an English translation!
My glossary workflow is:
- Scan the whole book or chapter with an AI agent that can create glossary items. These items should include not just source and translated terms, but also contextual notes and stylistic guidance. The in-story meaning of terms is important too, so that AI can interpret wordplay and references more easily.
- Review the items, discussing choices you doubt or don’t like with an AI that can search the web as well as your past documents.
- Attach the glossary to every translation, and use it to proofread and validate the result.
Translation memory (TM), on the other hand, searches for your past translations that are similar to the current passage.
TM is critical to ensure that similar terms, idioms and situations are translated in the same way across pages and chapters. It also gives both humans and AI a useful reference for style consistency.
In Transept, translation memory fills itself as you go: every paragraph you approve is remembered. Transept also remembers decisions:
- Previous versions and why they were rejected.
- Discussions with Literess and other humans.
- What AI thought when creating the current version.
- Text around the snippet for more context.
That way, the pun you rebuilt in chapter 3 comes back the same way in chapter 12, for the same reason. You can also scope the memory to a single book or share it across a whole series.
The same memory lets several AI agents translate one long book at once, sharing decisions as they go: that took a 40,000-word document from 8 hours to 50 minutes. More on how we built it in our article on translation memory.
4. Translate your story in smart segments, not paragraph by paragraph
If you translate paragraph by paragraph, the AI loses the scene. You also end up paying to resend your style guide and glossary on every single paragraph. Translating a scene or a few pages at a time keeps the prose connected and consistent, and costs a lot less.
What surprised us in our benchmarks is that bigger chunks did not make individual sentences better on their own. The real jump in word choice and phrasing came from telling the model to translate the meaning instead of the words, and giving it room to think.
That held across every model family and size we tested.
In Transept, we send around 1,200 words per call (never more than 1,500), a practical window we picked to keep scenes intact and costs down.
Make sure every call still gets the full context a human translator keeps open on their desk: the surrounding text and earlier approved lines from the chapter, your full style guide and glossary, and relevant translation memory matches (with the decisions behind them) from earlier chapters.
If you’re translating a comic or a layout-heavy PDF, send the whole page as one prompt together with its image. This way, the AI model can link the speech bubbles, artwork, and text together and understand the internal logic.
5. Use reviews to control the final translation
“Reading and checking after AI” is a truism, but it’s next to impossible to do it properly on a large text.
Research calls this post-editese: when a human edits raw machine output, the machine’s draft acts as a cognitive anchor. Across five translation directions, Antonio Toral (2019) found that post-edited translations mostly stayed closer to the source’s sentence lengths than translations from scratch (up to 38% closer), used a flatter vocabulary, and leaked source-language grammar. And in Guerberof-Arenas & Toral’s (2022) literary study, giving professional translators an MT draft to “fix” actively constrained their creativity compared to translating from a blank page. When a human reads too much machine, they start thinking in machine, too.
Luckily, you can use AI to check AI’s work!
Our internal benchmarks show that while AI makes plenty of mistakes on a first pass, it is much better at spotting them when asked to review with a clear rubric. Checking for drifted terms, literal idioms, and stiff phrasing is far easier for a model than getting them right on the first try. Taste calls, deliberate oddities, and humor that spans paragraphs still belong to you.
In Transept, we built the review flow so the AI does the first pass and you only step in where it actually matters:
- Smart Proofread checks the whole translation against the original and touches only the paragraphs with a real issue, such as drifted terms, missed glossary entries, or dropped lines. You can also give it a custom instruction (like “tighten the dialogue”) to flag passages that miss the mark.
- The review screen lets you confirm every edit by AI in one queue: proofread fixes, Literess’s edits, and open comments. Each card shows the original text, the before-and-after diff, and why the AI changed it, so you can accept, reject, or iterate with AI on any highlighted issue one by one.
- Review links let you send the draft to a native reader or editor so they can leave comments without creating an account.
Instead of reading through the whole book in raw machine output and going blind to its phrasing, you spend your energy only on the lines that need a human call.
Start with three pages
Most bad AI book translations fail in the first hour, not the last. People paste a whole chapter into a chat window, get flat prose back, and spend weeks fighting post-editese.
If you put your real effort into those first three pages instead, every step after that gets faster and stays closer to your voice.
Pick a 900-word slice of your manuscript and test the flow on it before committing to the whole book. If you want the style guide, glossary, decision-aware translation memory, and the review screen wired together in one place, you can run your first slice in Transept’s Author mode.
Sources
- Ferstler, Y. et al. (2026). AI translation of literary texts is “fine”, but readers still prefer human translations. Findings of the Association for Computational Linguistics: EMNLP 2026 (arXiv:2606.26040).
- Guerberof-Arenas, A., & Toral, A. (2022). Creativity in translation: machine translation as a constraint for literary texts. Translation Spaces, 11(2), 184–212 (arXiv:2204.05655).
- Toral, A. (2019). Post-editese: an Exacerbated Translationese. Proceedings of Machine Translation Summit XVII: Research Track, 273–281.
- Harry Potter Lexicon (2025). The Russian Philosopher’s Stone, January 27, 2025.
- Financial Times Magazine (2003). To the ends of the earth, May 17, 2003 (interview with Viktor Morozov).
- Outpost Dispatch (2003). Harry Potter in…Rusyn?!, Vol. 1, Issue 2, November 2003.
- A-BA-BA-HA-LA-MA-HA Publishers. The book series about Harry Potter.
- Ivakhnenko, M. (2026). Ten languages in three days: field notes from localizing an AI translation product. Transept Journal, July 12, 2026.
- Vlasiuk, V. (2026). Translation Memory: What It Is, How It Works, and Why It Matters for AI Localization. Transept Journal, June 30, 2026.
The author

Co-founder of Transept, writing as “Mevkh.” A Language and Literature degree, then a turn into software: senior AI engineer shipping production LLM features to 50,000+ users — RAG, agentic tools, LLM-as-judge evaluation. A novelist on the slow path, with 120,000 words of satirical romance fantasy in a drawer. The friction between AI translation and his own prose is what set this whole thing in motion.

