#Translate Novels With AI

AI can dramatically speed up fiction translation, but only when guided by a strong editorial workflow. Raw machine output of a novel is a draft, not a book: it reads sentence by sentence, while fiction succeeds or fails across chapters. This guide covers the failure modes specific to novels and the pipeline that addresses them.

#Why Fiction Is The Hard Case

Technical text tolerates literal translation; fiction does not. The recurring failure modes:

• Voice drift. A sardonic narrator in chapter 1 turns neutral by chapter 12, because each batch was translated without memory of earlier stylistic decisions.
• Name and title instability. "Lady Ashford" becomes "Mrs. Ashford" becomes a phonetic respelling. Honorifics are translated in one scene and preserved in the next.
• Literal dialogue. Machines translate what characters say, not what they mean. Idioms, sarcasm, and register (who addresses whom formally) flatten into correct-but-dead prose.
• Continuity leaks. A pun set up in one chunk pays off three chunks later; translated independently, the payoff dies.

None of these are fixed by a better model alone. They are fixed by workflow: constraints before translation, context during it, and human review after it.

#The Pipeline That Works

Glossary first -> Sample chapter -> Batch translation with refinement ->
Human review pass by pass -> Approve -> Export and verify

#1. Constrain before you translate

Build the glossary before the first run: character names, places, honorifics, invented terms, forms of address. In CatScribe this is a glossary group attached to the job — terms are shielded during machine translation and injected into the AI refinement stages, so your canon survives every pass. See Glossaries.

#2. Prove the setup on a sample

Translate one chapter end to end, review it, and export it before committing the book. A sample costs an hour; discovering a bad engine choice after 40 chapters costs a week. Use the sample review to harvest glossary terms you missed.

#3. Translate with refinement layers, in batches

A base machine translation pass followed by AI validation and rewrite stages produces noticeably better fiction than base translation alone. In CatScribe, Auto Mode packages this as quality presets — "Balanced" runs "Validation + Rewrite"; "Maximum" adds a technical check — using local AI models on your machine. Batching by chapter keeps runs recoverable and review manageable.

#4. Review like an editor, not a proofreader

Work in passes with a single concern each — the same discipline as human editorial workflow:

Quality scores (COMET-based, 0–100) are a triage signal, not a verdict: use them to decide where to spend attention. Approve chunks explicitly as you finish them — in CatScribe, approved chunks also feed a local translation memory that improves consistency in later batches.

#Character Voice: The Practical Method

Voice cannot live in a glossary, so externalize it:

1. Keep a one-page style sheet per book: narrator register, each major character's speech pattern, formality relationships between characters, how to handle dialect or verbal tics.
2. When using AI refinement on dialogue-heavy chunks, feed decisions from the style sheet as instructions ("keep Marta's speech clipped and informal").
3. Do a dialogue-only pass late in review: read only quoted speech, aloud if possible. Stilted dialogue is the most reader-visible AI artifact.

#What To Expect

Realistic expectations for AI-assisted novel translation:

• The machine draft gets you 70–90% of the way on prose mechanics, close to 0% of the way on voice decisions — those are yours.
• Review takes longer than translation. Budget accordingly; the translation run is the cheap part.
• Quality compounds across a series: your glossary, style sheet, and approved-chunk memory make book two faster and more consistent than book one.

For the screen-by-screen version of this workflow, see Translating Books.