When you upload a novel for translation, the process isn't just "send text to AI, get translation back." There are several stages, and understanding them helps you make better decisions about file prep and settings.
The system reads your file and extracts the translatable text. For EPUB files, this means navigating the package structure, identifying content documents, and separating text from markup. For TXT files, it's detecting encoding and identifying chapter boundaries from formatting cues like blank lines and heading patterns.
This is where file quality matters most. A well-structured EPUB with clean chapter breaks gives the parser clear boundaries. A messy file with broken references or mixed encodings forces the parser to guess, and guesses introduce errors.
AI models have context limits β they can only process a certain amount of text at once. The system splits your novel into chunks that fit within these limits while respecting natural boundaries (chapter breaks, paragraph endings, scene transitions). Good chunking preserves context. Bad chunking splits sentences mid-thought or separates dialogue from its attribution.
Before each chunk is sent for translation, the system checks your glossary for terms that appear in that chunk. Matching terms are included as translation instructions β "when you see η°δΈ, render it as Tanaka" β so the AI model follows your preferred terminology.
This is why a good glossary has such a large impact on output quality. Without it, the model makes its own decisions about names and terms, and those decisions won't always be consistent from one chunk to the next.
Each chunk is sent to the AI model with context about the target language, the glossary terms present, and any style preferences. The model generates the translation and returns it. Chunks are processed in sequence to maintain narrative flow where possible.
The translated chunks are stitched back together into the original file structure. For EPUB files, the translated text is placed back into the XHTML documents while preserving images, CSS, metadata, and navigation. For TXT, the chunks are concatenated with the original spacing and chapter breaks restored.
AI translation won't replace a skilled human translator, but it doesn't need to. For most readers who just want to enjoy a novel in their language, a clean file plus a solid glossary plus a quick editing pass gets you there. What used to take months now takes days.
