Transept: AI translation tool that remembers your style across projects

Transept AI is a translation workspace launched by writers Vitalii and Mariia, who built the product to solve a problem they faced directly. Vitalii had written 120,000 words of satirical romance fantasy—featuring mages drowning in debt and a multiverse economy—only to find that every translation tool he tried stripped the voice: tone, rhythm, wit, and emotional texture all flattened. Mariia had written semi-fictional stories about real people, mostly met and spoken to in English, and wondered what version of her subjects she missed by not writing in their native language. Neither existing translation tool could preserve the voice alongside accuracy.

To understand why, the pair benchmarked Claude, GPT, Gemini, and several open-source models against the same 50-page chapter, judging accuracy, tone, and whether satire still landed. None of them held all three; each failed in a different way. This led them to build the tool they wished had existed when they started.

Transept works in six steps. First, users bring in any document (DOCX, PDF, MD, TXT, or raw text) and get a side-by-side translation in their chosen language pair—translation is the starting point, not the destination. Second, users build a glossary once and toggle it on per project, so terminology stays consistent from page 1 to page 500 across all 12 supported languages. Third, and most distinctly, every translation a user approves is "remembered" by the system—not just the wording, but the reasoning: the alternatives rejected, the comments that explained the call. This "decision-context memory" is fed back to the model on the next run, across documents and across the team, so the AI learns not just the words but the judgment. Fourth, users can steer any sentence with a custom direction (e.g., "more formal," "keep the metaphor"), and the glossary and styleguide feed in automatically so the rewrite obeys terminology and tone. Fifth, a feature called Literess—described as a literary editor at the user's side—remembers projects and preferences between documents, reads the translation, flags where tone has flattened or meaning has drifted, and can act on the user's say-so: drafting documents, moving jobs across a board, or running workflows. Sixth, users review in a clean reading view with block-level comments, reply and resolve in place, and can send a link to a reviewer or client without emailing DOCX files.

Pricing is consumption-based and usage-only. The free plan includes 1,500 words per month with unlimited users and covers roughly 5 pages per month; it includes Standard mode (1 word per source word), translation memory, workflows, glossary, styleguide, QA, import and export, support for 50+ languages, and 10 lifetime messages with Literess. The Starter plan is €29/month for 100,000 words per month and adds Pro mode (3 words per source word for nuance), batch translation across whole projects, unlimited Literess, and free team creation with €29/month seats to fund shared word pools. The Pro plan is €79/month for 300,000 words per month and adds 3× Starter's allowance, priority support, and a predictable bill for most months without top-ups. Manual edits, exports, comments, and review activity are free; words are spent only when a block is translated by AI. Top-up word packs are available at 50,000 words for €19, 200,000 for €69, or 500,000 for €149, and pack words never expire. All prices are in EUR with VAT added at checkout where applicable. The service includes a 7-day money-back guarantee on the first subscription payment and allows cancellation anytime.

Transept emerged from a specific creative frustration: Vitalii translated 120,000 words of satirical fantasy and found that every tool—Claude, GPT, Gemini, and open models—preserved words but lost voice, with tone, rhythm, wit, and emotional texture flattening in the output. Mariia, a writer of semi-fictional stories about real people, faced the inverse problem: she had written in English to match the moments, but wondered what version of her subjects she missed by not writing them in their native language. Rather than accept the trade-off, the two read machine-translation research, benchmarked leading models on the same 50-page chapter, and found that none held accuracy, tone, and satire together; each failed differently.

The resulting product targets three distinct workflows—individual translators, content creators shipping in multiple languages before publishing windows close, and localization teams scaling across languages from a single source. The key innovation is "decision-context memory": instead of offering a static glossary lookup, Transept captures not just the approved wording but the reasoning behind each translation choice (the alternatives rejected, the comments explaining the call) and feeds that judgment back into the model on the next run, across documents and across team members. This approach aims to preserve not just terminology consistency but the translator's voice and intent.