From the editorial desk · English
How to Translate a Novel Without Losing Its Voice
Begin with what a scene does to a reader. A translation can preserve the events and still change who seems nervous, who is joking or who has power in the conversation.
The archive's review of Raak me. Niet. discusses how dialogue brings its characters to life. The review of Een jaar met jou gives conversations between Jamie and Ella an important place in the reading experience. Neither review is a technical translation assessment. Both explain why an author should test more than whether the plot survived.
Describe the voice before reviewing sentences
Write a short note for the narrator and each central character. Use behaviour you can point to: short answers when afraid, elaborate politeness when angry, a habit of avoiding someone's name. “Lyrical” and “natural” are too broad to settle a disagreement about a particular line.
Distinguish the author's intention from a preferred wording. “She makes a joke to avoid apologising” leaves a translator room to create a joke that works. Requiring the same idiom in every language may prevent that. Keep a separate list of names, invented terms, recurring objects and intentional ambiguities so the translator or editor knows where consistency matters.
Select scenes that can reveal a problem
An opening chapter may introduce the setting beautifully while containing little of the book's harder work. Choose a quiet exchange with subtext, a scene with several speakers, an emotionally heightened passage and something culturally specific. If the novel uses unusual syntax, include it. If a joke depends on a detail explained chapters later, tell the reviewer where that explanation lives.
For each passage, ask a reader of the target language to describe the scene without seeing your intention note. Who wants what? Which line is sincere? Who interrupts whom? Compare their answer with the original scene before discussing individual words. A mismatch does not automatically mean the translation is wrong; it tells you where to investigate.
What our public-domain excerpt translations showed
Method and disclosure. On 10 September 2026 we processed constructed excerpts from Lewis Carroll's Alice's Adventures in Wonderland and Jane Austen's Pride and Prejudice with the Skillsfor.ai application and its real translation backend. About my bookshelf and Skillsfor.ai have a common owner. The tests ran in an isolated QA environment, with test sign-in and no payment. These were three short file translations, not complete novels, a blind comparison or a native-language panel review.
The Alice EPUB was translated from English to Spanish and Dutch; the Austen DOCX was translated from English to Spanish. The application counted 3,960 characters for each Alice input and 3,069 for Austen. Some extracted Alice text included unwanted document titles and CSS, so these figures are application counts rather than clean literary word counts. The format study documents that problem separately.
| Passage | Observed output | What to investigate |
|---|---|---|
| Alice, English → Spanish; falling down the well | tuvo plenty de tiempo | The English word plenty remains inside a Spanish sentence. This is a concrete source-language remnant to correct, even though the paragraph continues fluently around it. |
| Alice, English → Dutch; the Rabbit's watch | het schoten haar te binnen | The verb form is inconsistent with the singular subject in this phrase. A target-language proofread must cover ordinary grammar as well as literary choices. |
Alice, Curiouser and curiouser!, English → Dutch | Nieuwsgieriger en nieuwsgieriger! | The source immediately comments on Alice's unusual English. Ask whether the target phrase and that following explanation still work together; translating the exclamation in isolation misses the joke's structure. |
| Austen, English → Spanish; the Bennets' dialogue | The output uses Spanish dialogue dashes and keeps separate speech turns. | A useful visible change, but it does not establish that irony, forms of address or the relationship between speakers are right throughout the passage. |
These examples are selected observations, not an error rate. We did not score every sentence, commission an independent human translation or establish which system is best. Download the input and output evidence package (ZIP) to read the surrounding passages. The files include awkward and incorrect language exactly as returned.
Use a review sheet that records effects
A useful comment names a location, describes the problem and explains why it matters. “Bad translation” does not tell an editor what to fix. “The reply sounds like agreement, but the character is avoiding the question” does.
| Check | Question for the reviewer | Record |
|---|---|---|
| Completeness | Are all speech turns, actions and negations present? | Source location and missing or added meaning |
| Speaker and register | Can you distinguish the speakers? Has formality, age or intimacy changed? | Speaker, relationship and line |
| Subtext | Does a joke, refusal or evasion have the same role in the scene? | Expected effect and observed effect |
| Rhythm | Where do sentence length and pauses affect pace? | A passage to read aloud, not an arbitrary word-count target |
| Consistency | Do names, recurring terms and habits remain recognisable? | Term or habit, first occurrence and later conflict |
| Ordinary language | Are there source-language remnants, agreement errors or unnatural combinations? | The exact phrase and a proposed correction |
Download the blank review sheet (CSV). Ask for reasons with suggested changes. Several plausible phrasings may preserve a scene; a single “correct” back-translation is not the goal. Back-translation can suggest places to inspect, but a second translation may smooth away the first one's error.
Carry the brief into the whole novel
After agreeing on a sample, maintain one glossary and one decision log. Record why a form of address or recurring phrase was chosen. Revisit the opening once later chapters reveal more about the characters. A word that seems incidental in chapter one may become a repeated motif.
Plan separate passes for source accuracy, target-language prose and proofing in the final reading file. Do not let a successful sample stand in for a full-book review. Our tests each used one translation chunk and cannot establish consistency across a long novel or a series.
For planning, formats and publication, return to Book Translation for Authors. For the experience of reading different editions, explore International Books.
Text sources: Project Gutenberg, Alice's Adventures in Wonderland and Pride and Prejudice, both listed as public domain in the USA. Our diagnostic files use English excerpts and our own test image, not a modern published translation or cover.