Multilingual Dictation: Two Languages, Auto-Detect, Mixing

You can dictate in two languages, but most built-in tools make you choose one at a time. Switching by hand takes a tap or a shortcut once both languages are set up. Automatic detection exists in several products, but it usually picks a single language for a whole recording or request. The hardest case is a sentence that mixes two languages. None of the vendor documents we read says its dictation follows a switch inside a sentence, and the research shows error rates rising sharply on that kind of speech.

This guide explains what each approach does, where the documented limits are, and how to set up a workflow that works around them. Terms are linked where they first appear: spoken language identification, code-switching and word error rate. For the wider subject, see our languages overview.

Most built-in dictation listens for one language at a time

Built-in dictation asks you to choose a language, then applies that language to everything you say. The setting belongs to the device or the keyboard, not to each sentence, so the choice you make at the start holds until you change it.

On a Mac, Apple says you can change the language you dictate in, or set up Dictation for several languages and switch between them as needed ¹. Apple's wording describes choosing a language when you need it. It does not describe recognizing two languages inside one sentence.

On an iPhone, Dictation follows the keyboard you are using. Apple says that when you use a keyboard for another language, you can use Dictation in that language as well ².

Chromebook dictation is stricter. It does not work if the language you speak does not match the device language ³, so a bilingual user has to match the device language to the language being spoken.

The table summarizes the cited statements above. The details differ by version, so check the settings on your own system before relying on them. For Mac specifics, see the Mac dictation languages guide. For Windows, see the Windows dictation languages guide.

Switching by hand takes one tap or shortcut once it is set up

Manual switching works well when each message or paragraph is in one language. The cost is one extra step each time you change language, which is small if you dictate in blocks. The steps differ by platform, and the exact menu names matter.

On a Mac, Apple's Mac Help describes this path for the language you dictate in: open Dictation in Keyboard settings, click the Edit button next to Languages, and select a language and region ⁴. The steps below follow that description.

1. Open Keyboard settings and select Dictation.
2. Click Edit next to Languages.
3. Select the language and region you want, and add the second language the same way.

On Windows 11, if the words you speak are not entered correctly, Microsoft says to check the voice typing language: press the Windows logo key and Spacebar, then choose one of your available languages ⁵.

On Android with Gboard, you switch languages in the voice toolbar by tapping the abbreviation at the bottom left, which shows your default language ⁶.

On an iPhone, changing keyboard changes the dictation language too, because Dictation follows the keyboard you are using ².

Automatic language detection usually picks one language per recording

Automatic detection is real, but it works at the level of a whole request, a recording or a long stretch of audio, not individual words. The documented tools ask you to list the candidate languages, and they then choose one. Two cloud services and the open-source Whisper family give the clearest rules.

Azure Speech takes a candidate list from your app. At-start mode accepts up to four candidate languages ⁷, and continuous mode accepts up to 10 ⁸. The service returns one of the candidates even if none of them is spoken ⁹, so a wrong or overly broad list can produce a wrong label.

Google's automatic language recognition accepts up to three languages ¹⁰. Google advises listing only the languages you need, because the fewer language codes you specify, the higher the likelihood that the service selects the correct one ¹¹. In a request with several languages, the service transcribes the audio in one best-fit language from the list ¹².

Whisper works the same way at file level. When no language is set, its transcribe function detects the language from up to the first 30 seconds of audio ¹³. The code then sets that language once and decodes the transcription with it.

The chart compares the candidate limits that the vendor documents state. It measures the size of the list each service accepts, not accuracy.

The pattern is clear. A larger list gives the service more room to pick a wrong language, and Google's guidance says the shorter list is the safer one. The question on whether dictation detects the language automatically covers the same settings.

Mixing two languages in one sentence is the hardest case for recognizers

Speech that switches languages inside a sentence is the case that the documented tools do not cover. Researchers call it code-switching. A 2025 study by Liu and colleagues says spontaneous switching within a sentence causes language confusion for code-switching recognizers, and that accent bias blurs the sound boundaries between languages ¹⁴.

The vendor documentation says the same thing from the product side. Azure's continuous mode does not detect a change of language inside a sentence. Microsoft's example is a speaker mainly using Spanish who inserts some English words: the service does not detect the language change per word ¹⁵. The Google documentation describes one best-fit language per request ¹². Neither document describes following a switch inside a sentence.

Measured error rates show the gap. In a 2025 study of two conversational code-switching corpora, Whisper's mixed error rate on code-switched utterances was about 39 percent ¹⁶. The same authors found that the speakers' original code-switching was altered during recognition ¹⁷. A 2025 Hinglish corpus paper cites recent evaluations that put the relative increase in word error rate for code-switched speech at 30 to 50 percent over monolingual input ¹⁸.

Coverage is also uneven. A 2025 review of code-switching speech recognition found that a small subset of languages receives most of the attention, and that the languages covered by datasets are very unequal ¹⁹. If your pair is Hinglish, Spanglish or another mix that is not widely studied, expect less support and more editing. The question on mixed-language recognition covers what that means for everyday use.

Multilingual models like Whisper can transcribe mixed speech but often lose the switching

Multilingual models are built to cover many languages, but they do not solve mixing. Whisper's training used a language token for each of 99 languages ²⁰, and OpenAI's speech to text guide says Whisper supports 98 languages ²¹. The two counts differ, so check the current list before you rely on a specific language. For the details by language, see the Whisper languages guide.

The studies show the limit. The 2023 authors of an adaptation method say that multilingual models such as Whisper often struggle with handling the code-switching setting ²². In a 2025 dataset paper on Mandarin-English dialogue, Whisper Large-V2 had the highest error rates among the pre-trained models tested, at 31.11 percent word error rate and 15.29 percent mixed error rate ²³. The 2025 voice anonymization study reports the same pattern on conversational speech, with the mixed error rate near 39 percent on code-switched utterances ¹⁶.

Training on mixed data narrows the gap. The 2023 method adapts Whisper to Mandarin-English code-switched speech, and it reaches a 14.2 percent mixed error rate while training only 5.6 percent additional parameters ²⁴. That number comes from a research experiment on a Mandarin-English speech corpus. It is not what a general-purpose Whisper app delivers on your own speech.

Accuracy also depends on the language, separate from mixing. The Voxtral paper's FLEURS Hindi table shows the spread across models. The chart uses the three Hindi figures in that table.

Whisper large-v3 had a word error rate of 28.87 percent in Hindi on that table ²⁵. In the same table, GPT-4o mini Transcribe had 12.36 percent ²⁶, and Gemini 2.5 Flash had 6.76 percent ²⁷. The three models are compared in one table, so the gaps are comparable within that test. A single table is not a verdict on any product, because versions change and test sets differ.

If you run Whisper yourself, set the language rather than letting it detect one, because detection uses only the opening window described above ¹³. For a recording that changes language, split it into sections and transcribe each one with its own language setting.

Names and borrowed words need a vocabulary list or spelling out

Names, brand words and loanwords are a common failure point for recognizers. A language setting alone does not fix them. A vocabulary list can help, when the tool offers one. Whisper's initial prompt can give the model custom vocabulary or proper nouns, which makes those words more likely to be transcribed correctly ²⁸. For steps by tool, see the question on dictating foreign names. For the general idea, see the glossary entry on custom vocabulary.

The built-in options vary. On a Mac, Voice Control accepts up to 1,000 custom vocabulary terms, and you can teach it how to say each one ²⁹. This is Voice Control, not Dictation, so check the language list before adding terms.

Third-party software gives more control over spelling. In Dragon, you add a word by specifying the word and its spoken form, and you can optionally train Dragon on its pronunciation ³⁰. For a name with an unusual spelling in either language, add it with its spoken form, and try that route first.

Hints help with a short list of words you use often, but they do not replace proofreading of a long mixed text.

Accents and diacritics are part of the spelling. If a name or loanword has an accent mark in its usual spelling, enter it with the accent in your vocabulary entry, so the transcript keeps it. Check that the tool keeps accents in the output you paste elsewhere.

Test a tool on your own mixed speech before you rely on it

Published numbers come from test sets that may not match your speech, so a short test with your own sentences tells you more. Write or record ten or so sentences you would really dictate, and include the mixed ones. Dictate them with the tool you plan to use, then count the errors.

Word error rate is the standard measure. It counts the words the recognizer substituted, deleted or inserted, and divides that total by the number of words you actually said. Multiply by 100 to get a percentage. The word error rate glossary entry gives the same formula with the normalization rules that change the score.

Run the test separately for each language and then for the mixed sentences. If the mixed sentences score far worse than the single-language ones, the tool is not following the switch, and you should plan to dictate in blocks. Ten sentences will not give a precise rate, but they will show whether a tool is in a usable range or far from it.

A daily workflow for people who write in two languages

The documented tools support a simple rule: choose one language per block of speech, keep the list of languages short, and proofread anything that mixes them. The steps below follow that rule.

1. Enable both languages on the device you dictate on most, using the menus described above.
2. Choose a language before each message or paragraph. Switch with the shortcut, the Globe key or the toolbar tap, and avoid switching inside a sentence.
3. Add names and recurring foreign words to the vocabulary list of the tool you use, with the spelling you want.
4. Use automatic detection only with a short list of languages you actually speak. For Google, that means no more than three.
5. Proofread mixed text. In the studies above, unadapted Whisper's mixed error rate on code-switched speech was 15.29 percent on one Mandarin-English test set ²³ and about 39 percent on conversational corpora ¹⁶, so expect to fix words that came out in the wrong language.

The flowchart shows the decision in order.

The table below turns the same logic into choices for common situations.

Two limits apply to everything above. First, we found no vendor document that reports error rates for mixed speech in Apple Dictation or Gboard, so for those products the advice rests on documented behavior rather than measured accuracy. Second, the product settings change with each operating system release, so the menu steps describe the versions we checked on 17 October 2026.

Yap, which we make, is one on-device dictation option for Mac and Windows. Check its language settings on the download page before you rely on it for mixed speech, and test it on a sample of your own sentences first.

Readers who arrive here with a specific question can start with the question pages: can you dictate in two languages at the same time, does dictation detect the language automatically, can speech to text handle Hinglish and other mixed languages and how to dictate foreign names and words correctly.

For related reading in this subject, see how to change the dictation language on any device, dictation languages by app and device, speech recognition and accents and Hindi voice typing and speech to text. For how to correct recognition errors in general, see the dictation mistakes guide.