English and Spanish. English and French. English and Portuguese, German, Italian, Ukrainian, Hindi. If you have ever dictated a sentence in your other language and watched it come back as nonsense, this page is about why that happens and what to do about it.
Speech recognition decides on a language before it decides on words. Almost every dictation tool makes that decision once, when the recording starts, and then holds it.
So you dictate in English, switch to Spanish for one sentence, and the software does one of two things. It writes that sentence out phonetically in English, which is gibberish. Or it translates it into English without telling you, which is worse, because it looks fine and is not what you said.
You end up doing what everyone does: dictating one language, stopping, opening a menu, changing a setting, dictating the other. Which is slower than typing, so eventually you stop bothering.
It asks the question once per sentence, not once per recording. The audio is split where you paused, because a pause is where a bilingual actually changes language, and each piece is identified on its own.
I'll send the draft tonight.
Pero primero hay que revisarlo todo.
Je te confirme demain matin.
One dictation · three languages · nothing configured
There is no setting to change and no language picker to remember. You hold a key, you talk, and the words arrive in whatever language you were speaking.
Because the obvious version does not work. Ask a model to choose freely between all ninety-nine of its languages and it is unreliable on anything short or noisy. On one second of my own recorded Ukrainian it scored Russian at 0.948 against Ukrainian at 0.032. Confidently wrong, and the wrong answer is the one that changes what your words mean.
Two things fix it. Restricting the choice to the languages you actually speak, because a language outranking yours does not matter if it is not a candidate. And adding probabilities together by language family before comparing — on a real microphone, Ukrainian often lands as Slovak, Czech and Polish at once, which is the model being certain about the family and unsure inside it. That change alone took a test set from four right out of seven to seven out of seven.
Those numbers came from my own microphone, in my own two languages. The method is not about Ukrainian: it works for whichever two, or three, you happen to speak, and there are twenty-nine that have been tested switching in and out of English and got every sentence right.
Every language on the list has been tested switching in and out of English and got every sentence right. Other tools advertise ninety-nine and let you find out which ones actually work.
Recognition happens on the machine. Audio is never written to disk and never transmitted. No account, no telemetry. Two things touch the network and I would rather say so than have you find them: a one-time model download the first time you use a language Apple's own engine does not cover, and a once-a-day version check that sends no identifier and can be switched off.
Fourteen days free. No card, no account, nothing to cancel. It is the same software you would be paying for.
macOS 26 · Apple silicon · signed and notarised · $39/year after the trial
Comparisons: against Wispr Flow · against MacWhisper