How to Translate a Book With AI Without Losing Its Voice

Lock a name glossary before chapter one, translate chapter by chapter, and give your bilingual reader a six-point checklist. Prompts included.

You can translate a book with AI and get a readable, publishable first draft, but only if you stop treating it like one big paste job. The method that works is chapter by chapter, with a locked name and term glossary written before chapter one, a short note describing the book's voice, and a bilingual reader who checks the result before anyone pays for it.

Skip those pieces and you get the usual problems. A character called Wren in chapter two becomes Reyezuelo in chapter nine. Banter turns polite. Every sentence comes out the same medium length. None of that is the model being bad at languages. It is the model being asked to make hundreds of small decisions with no memory of the ones it already made.

Here is the workflow I'd use for a full novel, whether you run it by hand in a chat app or in a tool built for it.

Can AI translate a whole novel well enough to publish?

Mostly yes, for plain narrative prose. Current models handle description, action and straightforward dialogue in major language pairs well enough that a careful editor can get it to a publishable state much faster than translating from scratch.

Where it still slips is predictable, which is good news, because predictable problems can be checked:

  • Names and invented terms drift between chapters, especially names that mean something (Thornwick, Ashfall, the Hollow King).
  • Register flattens. A rude teenager and a formal duchess end up sounding like the same polite narrator.
  • Forms of address wobble. In Spanish, German or French the model has to choose tú or usted, du or Sie, tu or vous for every relationship, and it will not choose consistently unless you tell it.
  • Wordplay, idioms and jokes get translated literally or quietly dropped.
  • Rhythm gets smoothed. Your clipped fragments become tidy complete sentences.

So treat the output as a strong draft, not a finished edition. That is also how Amazon sees it: KDP's content guidelines list translations created by an AI tool as AI-generated content you must disclose, even if you edit them heavily afterwards. More on publishing at the end.

What should you prepare before translating chapter one?

Most translation mistakes are decided before the first chapter runs. Spend an hour here and you save days of fixing.

  1. Pull every proper noun. Paste the manuscript into the Proper Noun Extractor and export the CSV. It runs in your browser and gives you the full list of character and place names, sorted, so nothing gets missed.
  2. Decide each name's fate. For every entry, choose: keep as is, adapt the spelling, or translate the meaning. Character names usually stay. Place names that are English words often should be translated, or readers in the target language lose the meaning you put there.
  3. Add invented terms and titles. Magic system words, ranks, honorifics, slang. Fantasy and LitRPG books need this most.
  4. Map forms of address. Write down who uses formal and informal address with whom, and the chapter where that changes, if it does. A shift from usted to tú can be a romance beat. Don't let the model make it by accident.
  5. Write a five-line voice note. POV, tense, the narrator's attitude, sentence habits, and anything the book does on purpose that looks like a mistake.

If you already keep a story bible, most of this exists. My guide on building a story bible AI models actually follow covers the format that models respect, and the same rules apply to a glossary: short entries, one decision per line, no essays.

To turn the raw name list into a glossary quickly, I hand it to a model with a prompt like this:

I'm translating my fantasy novel from English into Spanish (Spain). Below is the list of proper nouns from the manuscript, with a one-line note on each. For every entry, propose one rendering and mark it KEEP, ADAPT or TRANSLATE. Translate place names that are ordinary English words so the meaning survives; keep personal names unless they are clearly descriptive. Flag any choice that could clash with an existing real place or brand. Output a two-column table: English | Spanish. Do not translate anything else. [paste list]

This works because it asks for a decision per entry with a stated rule, so you can scan the table in five minutes and overrule the handful you disagree with, instead of discovering the choices one chapter at a time.

How do you translate chapter by chapter without names drifting?

The rule is simple: every chapter gets the glossary, the voice note, and a little of what came before. Then every chapter hands back anything new it had to decide.

  1. Send one chapter at a time. Long chapters can go in two or three parts, split at scene breaks.
  2. Include the glossary and voice note at the top of every request, not just the first one.
  3. Paste the last two or three paragraphs of the previous chapter's translation so the tone carries over.
  4. Ask the model to list any new names or terms it rendered, then add them to the glossary before the next chapter.
  5. Save each chapter's output as its own file. You will want to compare later.
Translate the chapter below from English into German. Follow the glossary exactly; if a name is in it, use that rendering every time. Forms of address: Mara and Jonas use du with each other; everyone uses Sie with Captain Hale. Voice note: close third person, past tense, short sentences in action scenes, dry humor, no added explanation. Keep paragraph breaks where they are. After the translation, list any name, title or invented term you had to render that is not in the glossary, with your choice. GLOSSARY: [paste] END OF PREVIOUS CHAPTER (German): [paste last 3 paragraphs] CHAPTER 7: [paste]

The closing instruction is the part people skip. It turns the model's silent choices into a list you can approve, which is how the glossary grows without drift.

If that sounds like a lot of copying, it is. On Entangled Text the AI book translator does this loop for you: you upload an EPUB, Word, PDF, RTF, ODT, HTML, TXT or Markdown file (or paste the text), pick any pair among 75 languages, and it translates chapter by chapter with a name and term glossary that carries every name forward. You can write your own glossary in the translation panel, and yours always wins. Translating a whole book is part of Pro ($15 a month) and Ultra ($25 a month), and the result opens as its own book so you can edit and export it.

If you'd rather do it in a chat app, pick a model with a long context window and strong output in your target language. The best AI models for writing page compares the current options for fiction.

How do you keep the author's voice in another language?

Voice is where a translation reads as a book or as a summary of one. The model's habit is to make everything clear, complete and medium length. Your job is to push back on that.

Three things carry most of a book's voice across languages:

  • Dialogue register. Each major character should keep their level of formality, their swearing, their verbal tics. If you built speech patterns on purpose, the guide on building a character's dialect and idiolect shows how to write them down as rules, and those rules go straight into your voice note.
  • Sentence length. Fragments should stay fragments where the target language allows it.
  • What the narrator refuses to explain. AI translators love adding a clarifying phrase. Subtext dies one helpful clause at a time.

I don't ask the model to "make it sound more like the original." That gets you a rewrite. I ask it to find the places where it drifted, and nothing else:

Below are an English passage and its French translation. Do not rewrite the translation. List the lines where the French changes the register (more formal, more polite, less blunt), merges or splits sentences in a way that changes rhythm, adds explanation not in the English, or drops a joke or idiom. For each, quote the line, say what changed in under 15 words, and suggest a minimal fix. ENGLISH: [paste] FRENCH: [paste]

Asking for a list of deviations with minimal fixes keeps the good 90 percent of the translation untouched, and gives a human reviewer a short list instead of a whole chapter to second-guess.

A quick warning about back-translation, where you translate the result back into English to check it. It catches dropped sentences and wrong facts. It does not catch stiff, unnatural phrasing, because a model will happily turn stiff German into smooth English. Use it for accuracy, not for voice.

What should a bilingual reader check before you publish?

You need a native or fluent reader of the target language. Not for every comma, but for the things models reliably get wrong. Give them a checklist so their time goes where it matters:

  1. Names in three places. Open the first chapter, a middle chapter and the last chapter. Every name and title should match the glossary.
  2. Dialogue read aloud. Ten minutes per chapter of reading conversations out loud finds stiff phrasing faster than silent reading.
  3. Calques. Expressions translated word for word that no native speaker would say.
  4. Forms of address against your map, especially in scenes where the relationship changes.
  5. Numbers, units and dates. Decide once whether miles become kilometres and stick to it.
  6. Cultural references that need a footnote, a swap, or nothing.

Before the reviewer starts, you can run a calque hunt yourself to make their pass shorter:

You are a native Brazilian Portuguese editor. Read this chapter and list every phrase that sounds translated from English: literal idioms, English word order, false friends, and dialogue no Brazilian adult would say out loud. Quote each phrase, give the natural Brazilian alternative, and stop. Do not comment on anything else. [paste chapter]

The narrow role and the "stop" instruction matter. Without them you get a general critique of the whole chapter, and the useful finds are buried in it.

When the reviewer's edits come back, run the raw AI chapter and the edited chapter through the Manuscript Diff tool. Seeing word-level changes shows you which kinds of mistakes keep coming up, and those patterns are worth adding to your voice note for the remaining chapters.

Publishing a translated novel: rights, disclosure and Kindle Translate

Three practical things before you upload.

Rights. If you self-published the original, you almost certainly hold the translation rights. If a publisher has your book, check the contract first. Translation rights are often licensed separately.

KDP disclosure. KDP's content guidelines define AI-generated content as text, images or translations created by an AI-based tool, and say it still counts as AI-generated after substantial edits. So an AI-drafted translation should be disclosed when you set up the title, even after a human review. Credit the translator properly in the contributor fields too.

Kindle Translate. Amazon runs its own AI translation service for KDP ebooks, but as of this month it is an invite-only beta covering English into French, German, Italian, Portuguese (Brazil) and Spanish, and those five languages into English. It won't take public domain books or books that are already translations. If your pair or your book falls outside that, you are doing it yourself anyway.

For the final cleanup and upload, the KDP upload checklist for AI-assisted drafts covers formatting and metadata, and the same steps apply to a translated edition.

If you only take one thing from this: build and lock the glossary before chapter one, and make every chapter report back the new names it had to decide. That single habit prevents most of what makes AI translations read like AI translations.

Try it on your own draft

Seven Amazon KDP keyword slots written from your manuscript. Each phrase stays under 50 characters, drops title words and banned terms, and explains reader intent. No invented search volume. Free, in your browser, no account needed.

Writing the whole book? Start it free — no credit card.

Comments (0)

Share a tip, question, or experience — keep it useful for other writers.

Prefer a public profile and edit access? Sign in or create a free account.

5,000 left

No comments yet. Be the first to share a thought.