# Do you have to set the meeting language before a notetaker joins?

> For anyone whose Google Meet or Microsoft Teams meetings are not always in one language: why many notetakers and caption tools ask for the spoken language up front, what a wrong setting does to a transcript, and what automatic detection has to get right — a meeting that changes language, a series that changes over time, and cost. Then how TrackHeros works without one.

Canonical: https://trackheros.com/blog/meeting-notetaker-auto-detect-language/  
Author: Sami Kanafani  
Updated: 2026-10-22

**In short:** Many AI notetakers ask for the meeting's spoken language up front, because speech models decode better when told and live captions are set per meeting; a wrong setting produces fluent text that was never said. TrackHeros needs no language setting: it works out each Google Meet and Microsoft Teams meeting's language from the meeting itself, from close to 100 it recognises, and writes the notes in the language you choose.

| Fact | Detail |
| --- | --- |
| Language setting in TrackHeros | None needed — detected per meeting; listing a client's languages is optional |
| Notes language | English by default, or eight other languages you choose |
| Spoken languages recognised | Close to 100 |

You do not have to set the meeting language if your notetaker can work it out, but many ask, and the reason matters, because a wrong setting does not produce an error — it produces a fluent transcript of things nobody said. This guide is for professionals whose Google Meet and Microsoft Teams meetings are not always in English, or not always in one language: why tools ask, what goes wrong when the answer is wrong, what automatic detection has to get right to replace the setting, and what to do meanwhile. The last section covers how TrackHeros handles it; the detail is in [automatic language detection](/features/automatic-language-detection/).

## Why do so many meeting tools ask for the language up front?

Because speech models decode better when they are told. Detecting the language is an extra step that can itself be wrong, so many tools hand the choice to you.

A speech recogniser does two jobs at once. It turns sound into candidate words, and it chooses between the candidates using what it expects the language to be. Told "Spanish", it has Spanish words to choose from, and the choice is easier. Not told, it has to decide what language it is hearing before it can decide what was said, and a short or noisy stretch of audio gives it little to go on.

Live captions add a second reason. On Google Meet and Microsoft Teams the captions are set to one spoken language per meeting. A notetaker that builds its transcript from those captions inherits whatever that setting was, and whoever chose it — the organiser, a participant, or nobody, leaving the default — made a guess. Why the platforms work this way is covered in [why meeting captions pick one language](/blog/why-meeting-captions-pick-one-language/).

None of this is unreasonable. If every meeting you hold is in Spanish, a setting that says Spanish works, and it is cheap.

## What goes wrong when the language setting is wrong?

You get fluent text that was never said. The model does not stop and does not warn you; it maps the speech onto the nearest words of the language it was told and returns tidy, punctuated sentences.

Set to English, a French aside about Thursday's deadline comes back as an English sentence about something else. Set to Spanish, a run of English technical terms comes back as Spanish-sounding words. In both cases the words that go wrong are the ones that matter, because they are the ones in the other language. Then the summary is written from that transcript, and it is just as fluent.

That is what makes a wrong language worse than bad audio. Garbled text gets checked. Fluent text gets forwarded. Signs that a transcript was decoded in the wrong language:

- Sentences that are grammatical but make no sense in context.
- Names, numbers and product terms that change spelling from one mention to the next.
- A long stretch where someone you know spoke at length reduced to a few short, odd lines.
- Action items that nobody remembers agreeing to.

The setting goes wrong in three predictable ways. The client set the meeting up in their language and you speak another. A recurring call that used to be in English gains someone new and changes language. Or the meeting is in two languages, and any single setting is wrong for half of it.

## What does automatic language detection have to get right?

Three things a fixed setting never had to handle: a meeting that changes language, a series that changes over time, and the cost of checking. Detection that gets these wrong is a setting with extra steps.

### A meeting that starts in one language and continues in another

The obvious design listens to the opening and fixes a language for the rest. It fails the same way a wrong setting does, just later. Meetings often open with small talk in a shared language and move to the substance in another; bilingual teams switch inside a sentence. Detection has to look past the first minute and accept that a meeting can be in two languages at once, which in practice means a model that detects the spoken language as it goes. [What happens when a meeting switches language mid-sentence](/blog/meeting-switches-language-mid-sentence/) covers that case in detail.

### A recurring call whose language changes

Remembering a series' language is sensible: the same people, meeting about the same thing, tend to speak the same way each week. But the memory goes stale. A new stakeholder joins, a team in another country takes over, a call that was in English gains a few minutes of German. A remembered language needs re-checking on a schedule, and one sighting of a new language should outweigh many quiet weeks.

### Cost

The reliable way to know what was said, in any language, is to transcribe the recording again after the call with a model that is not tied to one language. Doing that for every meeting is wasteful: an English meeting with clean captions gains nothing from it but the bill. A sensible design checks cheaply first and runs the expensive pass only when the cheap checks cannot rule it out. And when in doubt, it runs it, because being wrong toward the expensive pass costs money, while being wrong the other way costs the meeting.

## What should you do if your notetaker needs a language setting?

Set it as close to the meeting as the tool allows, and check the first transcript of any new meeting before you send anything built from it.

- **Set it per meeting or per series** if you can, rather than once for the account. One client's standup and another's review do not have to share a language.
- **Set it to the language the decisions are made in**, not the language of the greeting.
- **For bilingual meetings**, find out whether the tool accepts more than one language. If it does not, pick the one most decisions are in and treat everything said in the other as unrecorded.
- **Keep the recording.** A transcript you cannot check against the audio is one you have to trust.
- **Read numbers, names and action items against the recording** before a summary leaves your hands.

## How does TrackHeros handle the meeting language?

Nothing has to be set. Each meeting's language is worked out from the meeting itself, in stages, cheapest first: a recurring series' memory of the language it has been held in; a check of whether the platform captions read as coherent text; a short audio sample asked only which languages are spoken; then a probe and, when needed, a full audio pass over the recording. Every stage can send a meeting to the audio pass; none can keep it on the captions alone without independent confirmation. The decision is the same on Google Meet and Microsoft Teams. You can list the languages a client speaks, and that is optional; a language left off the list is treated as less likely, so if you list them, list all of them.

A meeting in another language, or in more than one, is transcribed by a speech model that detects the spoken language itself, then rendered into English line by line, each line keeping its speaker and timestamp. The summary, decisions and action items are written in English by default, or in Arabic, French, Spanish, German, Italian, Portuguese, Ukrainian or Indonesian, whichever you set for the account, the client, the series or the meeting. The transcript page shows one line — "Translated to English from a conversation held in French" — and that is the only place the transcript mentions language. Close to 100 spoken languages are recognised; [the languages page](/languages/) lists them all.

Recurring series remember their language. Once a series has been heard in English only, later occurrences can skip the audio pass; every tenth occurrence is checked afresh, and one sighting of a second language sends the series back to the audio pass at once.

If a transcript cannot be read reliably, the summary says so instead of inventing content. Switching a meeting's summary profile or notes language writes new notes, but there is no button to transcribe it again, so the recording is the reference — see [what to do when a notetaker detects the wrong language](/answers/notetaker-detects-the-wrong-language/).

## Not yet available

- A transcript kept in the language spoken, rather than rendered into English — not yet available

## Frequently asked

**Why does my notetaker ask which language the meeting is in?**

Because a speech model chooses between candidate words using the language it expects, and it decodes better when it is told. Live captions on Google Meet and Microsoft Teams are also set to one spoken language per meeting, so a notetaker built on them inherits that choice. Asking you is cheaper than detecting.

**What happens if the meeting language is set wrong?**

The model does not stop or warn you. It maps the speech onto the nearest words of the language it was told, and returns fluent, punctuated sentences that were never said. A summary of that reads just as confidently, which is why a wrong language setting is worse than a transcript that is visibly broken.

**Can automatic language detection handle a meeting that switches language?**

Only if it looks at the whole meeting and accepts more than one language. Detection that listens to the opening minute and then fixes one language for the rest fails exactly as a wrong setting does, just later. Mid-sentence switches need a model that detects the language as it goes.

**Does TrackHeros need a language setting?**

No. Each meeting's language is worked out in stages: a recurring series' memory, a check of the captions, a short audio sample, then a probe and a full pass over the recording when needed. The transcript comes out in English, the notes in the language you choose, and listing the languages a client speaks is optional; if you list them, list all of them.

**Is it wasteful to transcribe every meeting twice?**

Yes, which is why a careful design does not. An English meeting with clean captions gains nothing from a second pass but the cost. The sensible pattern is cheap checks first and the full pass whenever they cannot rule it out, with every doubt resolved toward the full pass.

---

TrackHeros is in early access: https://trackheros.com/#early-access