--- title: "What languages Omit actually works in" description: "Three different answers, because the interface, the name detection and the ID detection each cover a different set. Five interface languages, seven detection languages, and structured detectors that work in any language at all." canonical_url: https://omitsys.com/blog/what-languages-omit-works-in/ author: "Ranjan Singh" date: 2026-08-14 published: false --- ## In short The Omit interface is available in English, German, Spanish, French and Italian. Detection of names, organisations and places covers English, German, French, Spanish, Italian, Portuguese and Dutch. Structured identifiers such as passport and VAT numbers are matched by shape, so they are found in any language at all. About 45 of 394 interface strings are still English in all four non-English locales, clustered in the vault screens, the re-identify flow and the detection-mode hints. People ask whether Omit works in their language, and the honest reply is that the question has three answers. The app you look at, the part that recognises names, and the part that recognises ID numbers all cover different sets. Lumping them together is how a vendor ends up implying more than it delivers, so here they are separately. ## The interface: five languages, and not all of it yet The interface runs in five languages: English, German, Spanish, French and Italian. It opens in English and you change it in Settings, under Interface language. Menus, buttons, the review screen and the error messages are translated, and a test fails the build if a locale is missing a key the English one has. It is not finished, though, and you should have that from us rather than discover it. About 45 of the 394 interface strings are still English in all four of the other locales, which is roughly one string in nine. That is every multi-word string byte-identical to the English one in all five files, setting aside the product name, so you can count it yourself. They cluster in the vault screens, the re-identify flow, and the hints explaining the detection modes. If anything that number flatters us, because it ignores single words. Edit, Filter, Add, Done, Cancel, Passphrase and about a dozen more are also still English everywhere, and they are left out only because a one-word count is harder to define than a sentence count. The test guarding the translations compares which keys exist, not whether anyone translated the value, so none of this fails a build. Run Omit in German today and you will meet English text at those points. The count only goes down from here. ## Detection: seven languages Detection covers seven languages: English, German, French, Spanish, Italian, Portuguese and Dutch. Those are the languages where Omit loads a full language model and can recognise a person, an organisation or a place from context, not from a pattern. You may have noticed those two lists do not match. Detection has Portuguese and Dutch that the interface does not. That is deliberate rather than an oversight: the documents people need to redact are wider than the interfaces they are comfortable reading. A Dutch contract landing on the desk of someone who works in English is an ordinary Tuesday, and the tool should read the contract even when its owner never changes the menu language. ## Beyond those seven, patterns still work Now the part that surprises people. Outside those seven languages, Omit still finds a great deal, because most of what makes a document dangerous is not written in any language at all. An IBAN is an IBAN in every language on earth. So is a credit card number, an email address, a VAT number, a national insurance number, a passport number and a tax ID. Their format is the same wherever they appear, and the surrounding prose is irrelevant to them. Omit ships structured detectors for 31 countries, and they fire on a document in Finnish, Polish, Japanese or Arabic exactly as they do on one in English. We have measurement for this, not just assertion, and it comes from an unlikely place. When we benchmarked Omit against Google Cloud DLP, four of the languages in the test corpus ran without any language model loaded at all, on a bare tokeniser. Those cells still scored, because the entity types being tested were structured identifiers. That was an accident of how the harness was configured, and it turned out to be the cleanest demonstration we have that the structured layer stands on its own. ## Where the line actually falls So the shape of it is this. In seven languages you get everything: names, organisations and places from context, plus every structured identifier. In any other language you get the structured identifiers, and you do not get the model that spots that Kowalski is a surname. For a spreadsheet of account numbers that distinction does not matter at all. For a Japanese case file full of personal names it matters enormously, and you should assume we will miss those names. That last example is worth sitting with, because it is the sharpest edge of this. We ship structured detectors for Japan, Korea, China and Taiwan, so identifiers in those documents are found. We ship no language model for Japanese, Korean or Chinese, so personal names in those documents are not. Anyone handling documents in those languages should treat Omit as an identifier tool and not a name tool, and should not take our word for it either way without running their own files through it first. ## Which installer changes when, not which Which installer you choose changes when the languages arrive, not which ones exist. The offline installer carries all seven detection languages in the box and never touches the network. The light installer starts with English, ready immediately, and downloads any of the other six on request, once each. If the machine will never have a connection and the documents will not all be English, take the offline one. ## What comes next Where this goes next is more languages on the detection side rather than the interface side, because that is where the gap actually costs people something. We are not going to put a date on it. When a language ships, it will ship in a release note with the model named, the same way the current seven are. If your language is not on either list and you are weighing us up, the useful test is to run one of your own documents through the beta and look at what comes back. The structured identifiers will be found. Whether the names are is the question worth answering for yourself, and it takes about two minutes.