Speech to Text in French, German, Russian and More

Dictation works in the six major European languages covered here on both the Mac and Windows 11. The practical differences are the regional variant you pick and the errors that each language's grammar creates. French hides plural and verb endings in speech, so homophones and agreement slip. German joins words into long compounds that the recognizer has not seen before. Russian nouns take six cases ¹, so one word has many endings to choose from. Set the exact language and region, then dictate a few sentences of your own before you rely on the result.

The major European languages are supported on every platform

Each of the six languages is on the Windows 11 voice typing list and on Apple's Dictation list for the Mac. The lists are not the same, though. Windows often gives one entry per language, while Apple gives language and region pairs. The table shows how each platform names the variants.

Two platform rules matter as much as the list. Windows voice typing needs an internet connection ¹³. Microsoft's speech privacy page also says that voice typing converts speech to text using online speech recognition ¹⁴. On the Mac, the on-device list is shorter than the main list. Russian is on the on-device list ¹⁵, but Portuguese (Portugal) is not ¹⁶, so a language on the main list may not work on the device.

Whisper's published error rates for these languages fall between 4.0 and 8.3 percent

Support on a list does not tell you how accurate a language is. The open-source Whisper model has one language code for each of its languages. Its large-v2 model was tested on FLEURS, a benchmark of read sentences in 102 languages ¹⁷. Across the six languages in this guide, the word error rates fall between 4.0 and 8.3 percent ^{18 19}. English, the reference point, is 4.2 percent ²⁰.

The chart shows French at 8.3 percent ¹⁹ and Dutch at 6.7 percent ²¹. Italian is at 4.0 percent ¹⁸, Portuguese at 4.3 percent ²², German at 4.5 percent ²³ and Russian at 5.6 percent ²⁴. French is therefore about 1.8 times the German rate on this test ²⁵. Read these numbers as a ranking on one benchmark, not as a promise for your own speech. The paper's table gives one figure per language, and the Portuguese column does not separate Brazilian from European Portuguese.

French: homophones and agreement endings are where dictation slips

French dictation handles ordinary sentences well, but its errors cluster on words that sound identical. A 2008 study of French speech recognition looks at the pair et (and) and est (is). It notes that recognizers rely on language model weights to tell such homophones apart ²⁶. When the sounds match, the recognizer's guess about the surrounding words decides the spelling, so a wrong guess gives a real French word with the wrong meaning. Other pairs work the same way. The definition of homophone explains the term.

Agreement is the second weak point. Many French written endings, such as the plural -s on nouns and the verb endings after a subject, are not pronounced. The recognizer has to use context to choose the spelling, and it can pick the wrong form even when the sound is clear. Check every sentence for plurals and verb endings, and read the draft back before you send it.

French also has its own spoken punctuation words, which differ from English. Check the French command list before you rely on it. The guide to punctuation commands in other languages covers the commands for each language, and the question on French punctuation commands covers the French ones.

Windows lists French (Canada) and French (France) as separate entries ², so Canadian and European French are separate choices on that platform as well.

For French, the practical test is to dictate a paragraph that includes "et" and "est", plural nouns and past participles, then correct the errors once. That will show you the mistakes you are likely to repeat.

German: compound words and noun capitalization test the recognizer

German is the language where long words cause the most trouble. German builds compounds by joining words together, so one spoken word can be a long string that the recognizer's vocabulary does not contain. In a 2018 open-source German recognizer, the authors found that compounds unknown to the language model are quite frequent, because German is a very productive language ²⁷. The authors also count each such error as at least two errors: a substitution and an insertion error ²⁸. A single unknown compound therefore produces at least two errors in scoring, a substitution and an insertion, even though the speaker said one word.

This has a practical effect for you. A compound the recognizer does not know can come out as separate words, and scoring counts that as errors. Whisper large-v2 scores 4.5 percent for German on the FLEURS read-speech test ²³, but that test uses read sentences, so check compounds in your own writing.

German capitalization adds a second check. Nouns start with a capital letter, so a lowercase noun is an error even when the sound is right. Read the German text before you send it, and check the compounds in particular.

Swiss German is much harder, and no platform lists it as its own language

Swiss German is a family of dialects, and it is not the same as Standard German. The public Swiss German corpus STT4SG-350 holds 343 hours of speech from all dialect regions and was described as the largest public speech corpus for Swiss German ²⁹.

Dolev and colleagues tested Whisper on Swiss German audio in 2024. The large-v3 model, used without any training on Swiss German, had a word error rate of 23.0 percent on the STT4SG-350 test set ³⁰. A Wav2Vec2 XLS-R model that was fine-tuned on the corpus reached 14.0 percent on its test set ³¹. The authors conclude that Whisper can be used out of the box for Swiss German when the output you want is Standard German ³². They do not claim it writes dialect spelling.

A 2022 evaluation by the FHNW team compared its own Swiss German model with anonymized commercial systems. The authors reported that their model had almost half the word error rate of the best commercial system ³³. The FHNW team built that model, so treat the margin as their own measurement from 2022.

On the platforms, the Swiss variant appears only in some lists. Apple's on-device Dictation list includes German (Switzerland) ³⁴, and its auto-punctuation list does too ³⁵. Windows 11 voice typing lists German without a region ⁴, and Whisper has only one German code, "de" ³⁶. If you dictate in Swiss German, choose the Swiss German setting where one exists, expect to correct many words, and read the question on Swiss German speech recognition for the details.

Portuguese: choose Brazil or Portugal, because the two differ

Portuguese has two variants that matter for dictation, Brazilian and European. Apple's main Dictation list for the Mac includes both ⁶, and Windows 11 voice typing lists both as separate languages ⁵. Both platforms treat the variants as separate choices, so pick the one you write in.

The difference shows up on the on-device list. Apple's on-device Dictation for the Mac includes Portuguese (Brazil), but not Portuguese (Portugal) ¹⁶. If you dictate European Portuguese on a Mac, check the on-device list before you rely on on-device recognition.

Whisper does not separate the two. It has one code for Portuguese, "pt" ³⁶, and its published FLEURS score of 4.3 percent for Portuguese does not say which variant was tested ²². We did not find a published head-to-head word error rate for Brazilian and European Portuguese on the same system. For a direct answer on your own speech, read the question on Brazilian and European Portuguese dictation.

On the Mac, each variant is a language and region pair in the same list ⁶, and you switch between them with the steps in the final section.

Pick the variant that matches your spelling. Brazilian and European Portuguese differ in vocabulary and in spelling, so a text written in one variant will look wrong to a reader of the other. Set the variant before you dictate, not after.

Italian and Dutch: strong support with fewer regional options

Italian and Dutch have good support on both platforms, with fewer regional choices than French or Portuguese. Apple's main Dictation list includes Italian (Italy) and Italian (Switzerland) ⁸, and Dutch (Belgium) and Dutch (Netherlands) ¹⁰. Windows 11 voice typing gives each one entry: Italian (Italy) ⁷ and Dutch (Netherlands) ⁹. If you write Dutch for Belgium, the Mac list has a Belgian entry while the Windows list has only the Netherlands entry.

Accuracy is also good on the benchmark. Whisper large-v2 scored 4.0 percent for Italian ¹⁸ and 6.7 percent for Dutch ²¹ on FLEURS. Both figures come from read sentences, so expect more errors on casual or fast speech.

If you dictate on a Mac, check the auto-punctuation setting. Apple's documentation says that Mac Dictation adds some punctuation automatically in supported languages, and that you can switch this off with the Auto-punctuation setting in Keyboard settings ³⁷. The guide to Mac dictation languages explains where each setting lives.

Russian and other Cyrillic languages: good support, with case endings to check

Russian is on the Mac main list ¹², on its on-device list ¹⁵, and in Windows 11 voice typing without a region ¹¹. Accuracy on the FLEURS benchmark for Whisper large-v2 is 5.6 percent ²⁴, which is in the same range as German and Portuguese.

The errors can come from grammar as well as sound. A 2013 study of Russian speech recognition from the Karlsruhe Institute of Technology describes Russian as challenging because of its rich morphology ³⁸. Russian nouns can be declined in six cases, with singular and plural forms and three grammatical genders ¹. The recognizer must choose the right ending for each case, so a wrong ending is a wrong word even when the stem is right.

The same study shows the scale of the problem. A sub-word language model reduced the Russian out-of-vocabulary rate from 4.8 percent to 2.1 percent ³⁹. The 2013 system is old, so the numbers show why the method matters rather than how well current apps do.

Check the case endings in your text: nouns after prepositions, adjectives that agree with them, and verb forms that depend on the subject. For more, read the question on voice typing in Russian.

Smaller European languages have patchier support, especially offline

Beyond the six languages above, support thins out. NVIDIA says its Parakeet v3 model extends speech recognition from English to 25 European languages ⁴⁰, so an open-source tool can cover more of them. Whisper's training set covers 99 languages, each with its own language token ⁴¹. OpenAI's speech to text guide counts 98 languages that Whisper supports, and says that accuracy varies by language ⁴².

The gap is also about where the model runs. Windows voice typing needs the internet ¹³, and the Mac's on-device list is shorter than its main list ¹⁶. For a smaller language, check both the language list and the offline list before you buy or install a tool. A language that works online may not work without a connection, and a language that works in one app may not in another.

Test any smaller language with your own writing. Read a paragraph from your own work aloud, dictate it, and count the errors. That count tells you more than any list.

Pick the exact language and region, then test it on your own text

Four steps cover most cases. First, find your language and region on the platform's own list. Second, check whether the list you need is on-device or online. Third, dictate a paragraph that includes the hard words for your language, such as homophones in French, compounds in German, Portuguese variant spelling, and case endings in Russian. Fourth, correct the errors once and note the pattern.

On a Mac, adding a dictation language takes three steps. Open Dictation in Keyboard settings, click Edit next to Languages, and select a language and region ⁴³. With more than one dictation language set up, you can switch by clicking the language label next to the cursor or by pressing the Globe key and choosing a language ⁴⁴. Apple's names for these menus can change between macOS versions, so use the names on your own screen if they differ.

The overview of dictation languages links every guide on this topic. The guide to dictation languages compares the lists across Apple, Windows, Google and Whisper, and the guide to speech recognition and accents explains how accent and dialect can lower accuracy even for a supported language. If you mix two languages in one text, read the guide to multilingual dictation first. For a Mac, the Mac dictation language page shows how to add a language. For Windows, the Windows page covers the setup.

The figures here come from the sources linked in each fact and are dated October 2026. Vendors release new models and change their lists, so when a platform's own list differs from this page, trust the platform.