One language per meeting: preparing a multilingual call
Dictation in Dictera works out the language as you speak, and you can change language in the middle of a sentence. The AI Assistant does not. Each listening session runs in one language you choose. That is a decision rather than a gap, but it means a call in two languages needs a minute of thought before it starts.
Why the assistant is different from dictation
A dictation is short. You speak a sentence or a paragraph, the model works out what language it heard, and the next dictation is a fresh decision. Mixed-language dictation works the same way at a finer grain: each word is recognised and written in its own script. The languages on offer depend on which speech model you downloaded, up to 100+ spoken languages with the largest.
A meeting is an hour of continuous speech from several people, transcribed live on the Mac while the same Mac is drafting answers or waiting to write the notes. The assistant transcribes with Apple's on-device speech engine, set to one language for the session, and the languages on offer are the ones that engine supports on your Mac. Detecting a switch mid-sentence, reliably, in real time, across two speakers on one channel, would cost accuracy on every ordinary meeting to help the occasional bilingual one. So the assistant asks you to choose, and then it is very good at the language you chose.
Where you choose it
Assist and Meeting Notes each have their own language setting, in the panel and in Settings, so the standup you run in English and the customer interviews you run in German keep their own. The transcript is produced in that language, and so is the write-up: notes, minutes and chapters come out in the language of the session. Chat is the exception. It captures no audio and answers in whatever language you type.
Set it before you press Start. It is a setting for the session, not a switch to flip when the conversation changes language.
Assist answers in the same language
The rule covers answers as well as transcripts. Assist drafts its suggested answers in the session's language, so an interview conducted in German gets German suggestions, and a support call in English gets English ones. Each mode also chooses its own model, so a fast local model can serve Assist in one language while a larger one writes up meetings in another. What no mode will do is translate: a transcript is kept in the language it was spoken. If a colleague needs the write-up in another language, export it to Markdown, open the file, select the text, and Translate Selection shows it translated in a floating panel without touching the document.
A meeting in two languages
Choose the language decisions will be stated in. Usually that is the language most of the meeting is held in, but not always: a Spanish team presenting to an English client will spend most of the hour in English and settle the action items among themselves afterwards in Spanish. Which of those you need in the notes decides the setting.
What happens to the other language is worth knowing. The engine hears it as the language you chose, so a German sentence in an English session comes out as English-shaped nonsense, or as nothing. Those lines end up in the transcript as noise and carry little for the notes to work with. A decision made in the other language can be missed. The fix is a habit good multilingual meetings already have: restate the decision in the meeting's language. "So, to confirm in English: we ship on Friday, and Ana owns the release."
Names, terms and the background field
Proper nouns are where a fixed language hurts most. A French product name in an English session, a colleague called Zsófia, an internal codename: any speech engine will do its best, and its best will be wrong. Meeting Setup has a background field for exactly this. Put the product names, the project codenames and the people in it, and the write-up uses the spellings you gave rather than the ones the transcriber invented. If you connect your calendar, the invitee list fills the participants in for you. The setup belongs to each meeting, and a new one starts from your defaults, so the standup's vocabulary is typed once.
When dictation is the better tool
Two situations call for the other pipeline. The first is a one-on-one where you both switch language constantly and neither of you will restate anything. The second is your own notes during a mixed-language call: hold the dictation key, say what you want to remember in whatever language it occurs to you, and it lands in your notes app in the right script, punctuated. Translate as you speak goes one step further: say it in one language and have it typed in another, one of 29. The assistant writes up the meeting; dictation captures what you thought about it. They were built as two pipelines because they are two jobs.
Before a multilingual call
- Decide which language the notes should be in, and set it for the mode you are using, Assist or Meeting Notes.
- Fill in Meeting Setup: who is on the call, and every name and term the transcriber is likely to mangle, in the background field.
- Agree at the start that decisions and action items are stated in the session's language.
- Keep dictation ready for the notes you take for yourself, in whatever language they come.
If the same recurring meeting alternates languages from week to week, the language is the one thing to check before Start; the participants and the background in Meeting Setup carry over on their own, so the rest of the preparation is already done.
None of this takes more than a minute, and it is the same minute you would spend telling a human note-taker which language to write in. Everything else about how a meeting is captured is on the Meeting Notes page, and why nothing joins the call has its own article.
