Multilingual Dictation: How to Switch Languages Without Restarting

You are on a call with a client in Madrid. She explains her situation in Spanish, switches to English to reference a statute, then drops back into Spanish. You are dictating notes the whole time, but your dictation tool is set to English, so her Spanish quotes come through as nonsense. You stop, change the language, and start again. She keeps talking. You miss a detail.

Bilingual professionals know this problem well. Many dictation setups ask you to choose one language before you speak, as if people think in tidy monolingual blocks. Multilingual speech models make a different workflow possible: one model that can recognize several languages and detect which one you are speaking. Getting good results still depends on how you configure the model and how you speak.

This guide covers how language detection works in a model like Whisper, why some language pairs are harder than others, how to choose a model size, and the habits that make switching smoother.

Why Single-Language Dictation Breaks for Bilingual Speakers

The cost of manual switching is easy to underestimate because it shows up as small interruptions: stop speaking, change a setting, restart, and find your train of thought again. Each one is minor. Over a day of mixed-language work, they add up, and they pull your attention away from the conversation.

The deeper issue is that a recognizer set to one language interprets everything as that language. When a Spanish speaker drops an English technical term into a sentence, a Spanish-only setup may treat it as poorly pronounced Spanish and produce the wrong words. A multilingual model has learned many languages at once, so it can recognize the switch instead of forcing it into the wrong language.

Rigid interfaces compound the problem. If changing languages means stopping, picking from a menu, and starting again, the tool punishes natural speech. Auditory's homepage describes the alternative: choose from supported languages, use automatic detection, and keep mixed-language work in one dictation workflow.^[1]

How Multilingual Speech Models Detect Language

Understanding the mechanics helps you work with the model instead of against it.

OpenAI describes Whisper as "a multitasking model that can perform multilingual speech recognition, speech translation, and language identification."^[2] The Whisper paper explains that these tasks are represented as tokens the decoder predicts, so one model can handle steps that used to need separate components.^[3] Of Whisper's 680,000 hours of training audio, the paper says "117,000 hours cover 96 other languages" beyond English.^[3]

Whisper processes audio "with a sliding 30-second window."^[2] The reference implementation includes a `detect_language()` function that estimates which language a segment of audio contains before decoding it.^[2] That design has a practical consequence: detection is easiest when a stretch of audio is clearly in one language, and harder when the language changes in the middle of a sentence.

There is a useful distinction here between language identification and code-switching. Language identification answers "which language is this segment?" Code-switching is a change of language inside a segment or sentence. A model can be good at the first and still make mistakes on the second, because it has to handle both languages at once and find the exact point where they change. That difference explains much of the frustration bilingual users experience.

Your Language Pairs Matter

Not all language combinations behave the same way.

Languages that sound very different, such as English and Mandarin, or English and Arabic, give the model strong acoustic cues about which language is being spoken.

Languages from the same family, such as English and German, or Spanish and French, share more sounds, but usually keep enough distinctive features for the model to tell them apart.

Closely related languages, such as Spanish and Portuguese, Hindi and Urdu in casual speech, or Dutch and Afrikaans, share so much vocabulary and sound that the model can genuinely struggle to tell which one you mean, especially for shared words.

If you work with a closely related pair, test carefully. Speak a few minutes of real work content and count how often the output lands in the wrong language. If your tool lets you choose a fixed language for a session, compare that with automatic detection on the same passage.

The Right Model Size for Your Language Combination

Model choice is one of the biggest levers for multilingual quality. OpenAI notes that its English-only ".en" models tend to perform better for English, which also means they are the wrong choice when you need another language.^[2] For multilingual dictation, use a multilingual model.

Larger multilingual models generally handle more languages, accents, and switches more reliably, at the cost of more memory and a longer wait. OpenAI's table lists approximate memory needs from about 1 GB for base to about 10 GB for large in its reference implementation, and leaner runtimes need less.^[2]

The practical recommendation: if you switch languages often and accuracy matters for your work, test the largest multilingual model your Mac runs comfortably. If you mostly use one language and only occasionally switch, a smaller model may be enough. Measure on your own speech using the method in our guide to measuring word error rate.

Five Dictation Habits That Reduce Multilingual Errors

The model does the heavy lifting, but how you speak affects whether it succeeds.

Dictate in full clauses per language. Embedding a single foreign word inside an otherwise monolingual sentence is the hardest case for detection. "I need to file the *demanda* by Friday" asks the model to catch a one-word switch. Completing the thought in one language before switching gives it a cleaner boundary: "I need to file the lawsuit by Friday. *La demanda debe presentarse antes del viernes.*"

Pause briefly at language boundaries. A short breath between languages creates a clear audio boundary and keeps the sounds of one language from blending into the next.

Keep your accent natural for each language. Speak each language the way you normally do. A blended pronunciation that does not clearly belong to either language is harder for any model.

Start with the less dominant language. If you mostly speak English but switch to Mandarin occasionally, begin a session with a Mandarin sentence or two so the model encounters both languages early.

Use each language's pronunciation for proper nouns. When dictating in German and mentioning München, say "München" the German way rather than "Munich" in English. Pronunciation cues help the model choose which language to write.

Example Setup: English-Spanish Legal Notes

Here is an illustrative configuration for a bilingual immigration attorney who dictates client notes, switching between English legal references and Spanish client statements.

The Configuration

A large multilingual Whisper model runs locally on the attorney's Mac. Automatic language detection is enabled. Filler removal covers both languages, including Spanish fillers such as "este" and "o sea." Grammar cleanup respects Spanish punctuation, including inverted question and exclamation marks.

Before and After

Here is an illustrative example of raw bilingual dictation:

```

the client states that um she arrived in 2019 bajo un visa de turista

and overstayed because este su esposo was deported and she had no means

of returning the relevant case is matter of rodriguez which establishes that

```

After language detection, filler removal, and cleanup:

```

The client states that she arrived in 2019 bajo una visa de turista

and overstayed because su esposo was deported and she had no means of

returning. The relevant case is Matter of Rodriguez, which establishes that...

```

Notice what still needs human review: the case citation, the date, and whether the Spanish grammar fix ("una visa") matches what the client said. Legal names and citations should always be checked against the source.

How to Measure Your Own Results

Over a few weeks, track:

• Word error rate on segments in each language
• Errors at mid-sentence switches
• How often the output lands in the wrong language

Then adopt the clause-boundary habit and measure again. Your own before-and-after numbers will show which change helped most.

Privacy: Cloud Multilingual Services vs On-Device Models

Major cloud speech services support multilingual recognition and language detection. The question for many professionals is not accuracy. It is where the audio goes.

With a cloud service, dictated speech goes to the provider for processing. For a bilingual attorney dictating client statements, that audio may contain privileged information, and possibly health information or personal data from clients in other countries. Each of those brings its own obligations. Under the GDPR, for example, a processor "shall not engage another processor without prior specific or general written authorisation of the controller,"^[4] and under HIPAA a vendor that receives protected health information on a covered entity's behalf is a business associate.^[5]

With an on-device Whisper configuration, transcription runs on your Mac, so no speech vendor receives the audio for that step. Review downstream apps, backups, and document systems as separate parts of the workflow. That distinction matters for professionals whose clients span several jurisdictions.

Latency depends on your hardware, model, and network. Local processing removes the network round trip; cloud processing removes the load from your Mac. Test both on the same recording if speed is your deciding factor.

Your 30-Minute Multilingual Dictation Setup Checklist

1. 1.Check your Mac (2 minutes). Note your chip and memory, and open Activity Monitor to see how much memory is free with your usual apps open.

1. 1.Choose a model (3 minutes). Use the table above. Pick a multilingual model, not an English-only one, and avoid turbo if you need translation.

1. 1.Enable automatic language detection (1 minute). Switch from a fixed language to automatic detection in your dictation tool's settings.

1. 1.Run a short calibration test (5 minutes). Read a passage that switches languages at each sentence, using your real language pair. Count the transitions the model got wrong.

1. 1.Practice the clause-boundary habit (15 minutes). Dictate real work for fifteen minutes, finishing each thought in one language before switching. Review the output and note where detection failed.

1. 1.Track your switch error rate (ongoing). Each week, divide misidentified transitions by total transitions in a representative session, and watch the trend.

If you are still choosing software, our roundup of Mac dictation apps with on-device processing and multilingual support compares the options.

Remember the call with your Madrid client? With a suitable model, automatic detection, and a few habits, Spanish can flow into English and back with fewer interruptions, and the transcription step stays on your Mac.

References

[1] Auditory, On-device speech to text for Mac. https://auditoryapp.com/

[2] OpenAI, Whisper repository README. https://github.com/openai/whisper

[3] Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision (2022). https://arxiv.org/abs/2212.04356

[4] GDPR Article 28, Processor. https://gdpr-info.eu/art-28-gdpr/

[5] 45 CFR 160.103, Definitions (Cornell Legal Information Institute). https://www.law.cornell.edu/cfr/text/45/160.103