Own Words

Ta strona jest po angielsku.

How Own Words works

Own Words drafts the next paragraph of your translation the way you would have written it. It learns your voice from anything you have written: past translations, your own books, articles, even transcripts. Have nothing? It learns from every correction you make. Word in, Word out, with your formatting.

There is a large language model in the middle, and we say so. What we built is everything around it: what it reads before it writes, which model does the writing, and how we check whether any of it actually helps. We test on real novels, with a working literary translator.

The first ten pages are free, so you can judge it on your own book.

It reads what you've written

Drop in anything you have written in your language, as Word or PDF, in whatever state the publisher left it: past translations with or without their originals, your own books, articles, transcripts. Own Words works out which language each file is in on our own server, in any of 110 languages, sorts a pile of files into "this is the translation of that" by itself, and keeps everything for next time.

Where a translation has its original, it pairs every paragraph of yours with the paragraph it translates. That sounds like bookkeeping. It is the part that decides whether anything after it works, because you don't translate paragraph for paragraph. You split one in two because your language breathes differently. You merge two short ones. Now and then a line is left out, by you or by an editor.

A pairing that goes by length is right until your first split, and wrong for every paragraph after it. Ours goes by what the paragraphs mean. It survives splits, merges and omissions, works in any alphabet, and flags the stretches it isn't sure of instead of guessing. Flagged pairs never become examples in your drafts.

It remembers how you solved it last time

Before drafting a paragraph, Own Words looks through your finished books for the places where you solved something like it: the same kind of banter, the same kind of landscape, the same awkward idiom. It picks a handful that show different angles rather than ten copies of the closest match, and skips the one-liners, because "He said nothing." is not an example of anything.

Alongside those it sees the paragraphs you just translated, notes on how you translate (written once from your finished book, and yours to edit), your glossary, who is who in the new book, and your corrections.

It drafts in your voice

From all of that, a language model writes the next paragraph. Three alternatives are one click away, and you can switch models whenever you like.

We did not pick the model by reputation. We ran more than a dozen of them on a translator's own novel, with the same instructions and the same paragraphs, and scored each draft by how much of it she would change. The answer key was work she had done before the tool existed, and no model could see it.

The famous, expensive models were not measurably closer to her than the one we use by default. Several were measurably further away. The very cheapest made mistakes of meaning, not taste: one turned stubble into a staircase. So the default is a modest model, and that is a finding, not a compromise. If a model refuses a scene (crime novels have this problem), another one takes over.

It learns from your corrections

When you change a draft, Own Words records exactly what you changed: this word for that one, that phrase moved to the start. Every change goes into the drafts after it, chosen for the paragraphs most like the one you are on. When the same change shows up twice, it becomes a rule you can see, and strike out if it's wrong. Rules are counted from your edits, never invented by a model.

Started with no files at all? This is how it learns your voice anyway: every correction goes in, repeated ones become rules, and we keep measuring, on real novels, whether each change brings the drafts closer to the translator. Changes that don't, we drop.

How we check

Every comparison runs on paragraphs the model cannot have seen the answer to. A difference only counts if it is bigger than chance: the same paragraph drafted twice already differs more than most "improvements" do. And on real work you sometimes get two drafts side by side, unnamed, in random order, and pick one. A change only wins when translators prefer it clearly and often.

What we tried and dropped

We tried several ideas that sounded clever. A fancier search for examples, and rules about the translator written by a model, did not bring the drafts any closer to her. The rules were even true about her, and still pushed every paragraph toward them. We don't sell any of it. The upgrade we didn't ship tells the story.

Want to see it on your own book? The first ten pages are free.