Last updated: 2026-10-16
To translate a whole book with an LLM and receive a usable EPUB, someone must handle more than the translation prompt. The source needs parsing, passages need tracking, terms need context, suspicious output needs checking, and the text needs packaging back into a book. A paid whole-book workflow takes responsibility for those steps, while still leaving limits that a reader should understand before buying.
If you started by considering ChatGPT, Claude or GPT-4, the useful question is what surrounds the model: how does the workflow account for the whole source and deliver the requested file? A model name alone does not establish translation quality, completeness or EPUB delivery. This guide explains the managed workflow and its evidence; it is not a model ranking or an API programming tutorial.
EPUB Translator accepts EPUB and text-based PDF books, charges once per book without a subscription, and includes both a translation-only EPUB and a bilingual EPUB. The bilingual edition places the original above the translation paragraph by paragraph. You do not need to supply API credentials, prompts or a model choice. Check current pricing: $0.004–$0.0125 per 1,000 source characters, minimum $3.99 per book. Your file's exact quote comes after sign-in and server parsing.
A chat model may accept a large amount of text, and custom tooling may generate an EPUB. Neither capability is impossible. Context limits vary by model and interface, so a fixed token-window comparison would be a poor buying rule.
Three separate questions still need answers:
Ordinary chat produces text; an EPUB-aware tool or a person must supply the book assembly and checks. Pasting text alone does not carry an EPUB's manifest, reading order, stylesheet relationships and internal links along with it. A custom workflow can manage those things, but its delivery should be evaluated separately from a model's prose.
An EPUB contains HTML, styles, images and metadata describing its resources and reading order. The managed workflow extracts translatable text while retaining where it belongs, then reconstructs the book around that mapping. That is why chapter boundaries, headings, links and inline emphasis need attention alongside the words.
Text-based PDF input adds a reconstruction step: a page-oriented document becomes an EPUB. It is not a promise of identical PDF page breaks or a translated PDF download. Scans and image-only PDFs are unsupported because this service has no OCR. DRM-protected files are unsupported too.
Names, technical terms and forms of address recur across a book. Passage context and terminology guidance help the model handle them consistently. They do not guarantee that every name, pronoun or term will be correct or identical across chapters. Instructions can be followed imperfectly, and a consistent translation can still be consistently wrong.
For a reader, the bilingual edition is a way to inspect passages after delivery. It is not a substitute for understanding the source or arranging human review when meaning matters. No human proofreading or in-app editor is included.
A workflow needs to distinguish “text returned” from “text translated.” EPUB Translator checks eligible prose passages for source-language carryover using a local completeness check alongside Jev, a decision model used here to assess whether translation happened. Jev does not generate the book's translation or certify its literary quality.
A flag from either check triggers retranslation. Persistent problems can escalate to a stronger translation model. Short passages and non-prose content are not all eligible for the same checks. When Jev is unavailable, the local check continues with narrower capability.
These checks address one failure class: text left in the source language. They do not score meaning, tone, terminology or every kind of omission. A completely target-language paragraph can still mistranslate the source.
Our public untranslated-passage report describes a 2026-10-11 German-to-Polish incident containing 2,803 eligible prose passages. The checks disagreed on 16 passages, all of which were manually reviewed. That review identified a known set of 37 untranslated passages across the checks' results.
| Check | Known untranslated passages detected | False alarms in that incident |
|---|---|---|
| Jev alone | 25/37 | 3 |
| Local check alone | 36/37 | 0 |
| Either check flags | 37/37 | 3 |
The disagreement review explains why the two checks are used together. Jev passed 12 passages that were mostly German with only one or two nouns changed to Polish; the local check caught them. The local check missed one short half-translated passage with too little comparable text; Jev caught it. The three false alarms concerned text that had been translated, even though it contained wording or translation problems. Those were false positives for untranslated-text detection, not evidence of successful detection of that failure class.
This was not an exhaustive or blinded audit. It investigated disagreements in a book selected because of a known problem. Both checks may have missed other passages. “37/37” describes detection of the known set; it is not accuracy, a zero-omission guarantee, a final-output score or proof of translation fidelity. It does not rank translation models or validate every language pair.
Two separate healthy books contained 573 English-to-Portuguese and 591 English-to-Czech passages. Jev, the local check and their combination each produced zero false alarms in those 1,164 passages only. This does not establish a general false-positive rate, and it does not cancel the three false alarms in the problematic German-to-Polish book.
Finding a problem and fixing it are separate outcomes. If retries or escalation fail, or escalation is unavailable, delivery can continue with a problematic passage from the original translation model, and the issue is logged. A completed download does not establish that every final passage was translated.
This is the practical value and limit of paying for checks and retries: there is a process for responding to suspicious passages, but there is no promise of a flawless book. The report contains aggregate results; this page does not offer private excerpts, a finished-book sample or a screenshot as evidence of final quality.
The following describes current packaging behavior. It is documentation of the deliverable, not a rendered sample or a claim that every source layout was visually verified.
| Included behavior | What it means for your book |
|---|---|
| Two EPUB editions | Translation-only and bilingual EPUBs are both included for every book, with no pre-translation output-mode choice. |
| Interleaved bilingual text | Matching prose blocks put the original above the translation paragraph by paragraph. They are not parallel columns. |
| Safe pairing limits | Identical text blocks are not duplicated. When a chapter's blocks cannot align safely, that chapter falls back to translation-only text. |
| Layout for the target | For targets needing horizontal text, vertical styles are adjusted and ruby annotations are removed from translated text; inserted original blocks retain their annotations. Japanese and Chinese targets retain source vertical-layout settings. |
| Text PDF handling | The source is rebuilt into EPUB. Output is EPUB only, with no promise of exact PDF pagination. |
Translated text is assembled within the source EPUB structure, but text length, fonts and reader behavior can change the visible layout. Check the table of contents, chapter order, footnote links, images, tables and any complex content in the reader you actually use. A well-formed EPUB is useful evidence about packaging, not proof of accurate prose.
The complete workflow is select file → sign in with Google or email → server parses and shows the exact file quote → Start translation → explicitly select target language → checkout/payment → automatic translation → download.
First select the book at the upload entrance. After sign-in, upload and server parsing produce a quote from that file's source-character count. A book of about 500,000 characters costs about $4.75; use your own file's quote to decide. This is a one-time book purchase with no subscription, not an API bill you must calculate from experimental requests.
Clicking Start translation opens the language picker. Opening and closing it does not create an order, spend balance or launch a job. The target is not preselected from the interface language or an earlier preference; you make an explicit choice for this book.
If payment is needed, choosing the language takes you to checkout. Review the book title or filename, selected target and total before paying; you can return to change the language. Successful payment starts translation automatically. If existing balance covers the book, an explicit notice before selection says that choosing a language will spend balance and start translation; that branch uses the balance without a separate checkout. A changed balance requires an updated notice and another explicit selection rather than a surprise deduction.
After completion, download either included EPUB edition or both. You are paying for the book workflow and files, rather than for access to a chat session or a promise of human-reviewed prose.
Start with passages from the beginning, middle and end, including chapter and paragraph endings. The completeness report shows why checking only an opening page is insufficient. Read whole sentences around unchanged words: a proper name can legitimately stay unchanged while surrounding source-language grammar signals a different problem.
Use the bilingual EPUB to compare those passages if you can read the source, or ask a bilingual reader to help. Separately check names, recurring terms, numbers and any passages where a misunderstanding would matter. Untranslated-text detection does not perform those judgments for you.
Open the EPUB in your intended reader and follow the table of contents and footnotes. Inspect tables or formulas your reading depends on. There is no in-app editor; adjustments can be made with separate EPUB software such as Sigil or Calibre. Publication or other accuracy-sensitive use needs independent human review.
For a choice between hosted, desktop and document workflows, the seven-tool comparison compares documented deliverables and setup. If a paid whole-book translation with both EPUB editions fits your needs, check pricing, then select your book to sign in and review its exact quote.