Speech to text in seven languages, on your own computer

The seven languages, and how well each one does

We measured every language on the same test before offering it, and we publish the results. "Words wrong" is the share of words the engine got wrong on 80 sentences read aloud, taken from Google's FLEURS dataset; the time is the median it took to decode one of them on the processor of a 12-core desktop PC running Windows. The method, in full, is on the evidence page.

Three things to read into those numbers, and one not to. The Spanish figure is Latin American Spanish and the Portuguese figure is Brazilian Portuguese, because those are the dataset's splits; no regional accent was measured, so nothing here speaks for Spain's Spanish, European Portuguese, Swiss German, Flemish, or Quebec and Belgian French. The sentences were read aloud, which is easier than the way most people dictate. And the times are for one fast desktop; a laptop is slower. What not to read into them is a comparison with other apps: they measure Deixis's engines, and the evidence page compares them with the Whisper model Deixis used before.

A language is offered only when an engine gets fewer than one word in ten wrong on this test. We measured Afrikaans, Zulu and Xhosa as well, and none came close, so they are not offered.

One page per language, with what Windows and others offer in it:

• German speech to text: 7.4% words wrong.
• Dutch speech to text: 9.2%.
• French speech to text: 9.2%.
• Italian speech to text: 3.4%.
• Spanish speech to text: 3.1%, Latin American Spanish.
• Portuguese speech to text: 5.6%, Brazilian Portuguese.

How the languages work in Deixis

You choose the language, one at a time. The first time Deixis runs, it proposes the language your computer's display is set to, and you can change it in settings whenever you like. It does not detect the language from your voice, and dictating in two languages inside one sentence was not measured, so we make no claim for it.

Two engines, both on your computer. English uses Whisper, as it always has. The other six use NVIDIA's Parakeet v3, a speech model that runs on your own processor. The first time you pick one of the six, Deixis downloads it once, about 670 MB. After that, dictation in that language works with the Wi-Fi off, and once the speech model is on it, the audio of a dictation never leaves the machine, under any setting.

What 1.6 promises in the six new languages is push-to-talk. Hold the chord, speak, let go, and the text is typed at the cursor in any app that takes text, with your clipboard handed back the way you left it. That is the part we have measured and checked, and it is what these pages claim.

The polish around it is tuned for English. Cleanup, which tidies a sentence after it is recognised; the dictionary, where you teach Deixis your own words; the five vocabulary packs, which are English word lists; and the punctuation Deixis adjusts to the field you are typing into: all four are built for English, and we do not claim they work in the other languages. Parakeet also hears your dictionary words only by their sound, so unusual names come out right less often than they do with Whisper in English.

Where it runs

Deixis 1.6 dictates in all seven languages on two platforms:

• Windows 10 and 11, 64-bit, on a processor with AVX2: Intel Haswell or AMD Excavator, 2013 and later. On an older processor it stops with an error rather than running slowly. The chord is Ctrl+Win.
• macOS 13 or later on Apple silicon. The chord is Ctrl+Option.

Android follows later, and we do not give a date for it. The figures on this page were measured on Windows, and none of them were measured on a Mac.

Windows 11 voice typing languages

Windows 11 has voice typing built in: press Win+H in a text box. Microsoft's support page, read on 4 November 2026, lists 46 languages for it, among them German, Dutch, French, Italian, Spanish (Spain and Mexico) and Portuguese (Brazil and Portugal). The same page says voice typing uses online speech recognition powered by Azure Speech services and that you need to be connected to the internet. No Microsoft page we read says it works offline in any language. The one on-device exception on that page is fluid dictation, a Copilot+ PC feature for English only.

Windows has a second, offline option, and it deserves the credit. Voice access, on Windows 11 22H2 and later, recognises speech on the PC after a one-time language download. Its set-up page lists English, Spanish (Spain and Mexico), German, French (France and Canada), Italian, Chinese and Japanese. It has no Dutch and no Portuguese. It is built for controlling the computer by voice, with dictation as part of that, and it is free. The older Windows Speech Recognition is deprecated: Microsoft says voice access replaced it on Windows 11 22H2 and later in September 2024, and its languages were English, French, German, Japanese, Chinese and Spanish.

Windows voice typing compared with Deixis goes through the rest.

Wispr Flow, Superwhisper and others

Wispr Flow. Our facts file on Wispr Flow, checked on 7 October 2026, records no language list, so we do not print one here; its own help pages are the place to read it. What its pages do say is that every dictation is transcribed in its cloud and that an account is required, so in whatever language you speak, the audio leaves your computer. Wispr Flow compared with Deixis.

Superwhisper. Our facts file records no language list for it either. Its own pages, read on 27 September 2026, say the free plan runs on local Whisper models, and that its Parakeet models and its cloud voice models are part of Pro. Superwhisper compared with Deixis.

Murmure, a free, open-source app, runs a tuned version of Parakeet v3, the same model family Deixis 1.6 uses, on Windows, Mac and Linux. By its own pages, read on 4 November 2026, it covers 25 European languages and detects the language from your voice. If you need Linux or automatic language detection, it is a fair choice.

Apple Dictation on the Mac recognises German, Dutch, French, Italian, Spanish and Brazilian Portuguese on the device once the speech model is downloaded, per Apple's feature-availability page (4 November 2026); European Portuguese is not on that list. Apple Dictation compared with Deixis.

Free, and where free stops

Dictation is free in every language, forever, with no account and no trial clock: everything that runs on your own computer is free. What can cost money is the part that needs a language model: cleanup by a model rather than by rules, translating a selection, and meeting notes. Those are free with your own provider key or a model on your own machine; otherwise Deixis Pro runs them for you, at $49 a year or $10 a month. Sharing a file as a link needs a Deixis account.