An AI notetaker that works out the meeting language on its own

TrackHeros is an AI meeting notetaker that connects every Google and Microsoft calendar you hold, recognises close to 100 spoken languages on its own, writes the notes in nine of them, and keeps every client's notes in a sealed workspace of its own. This page is about the middle part: how the language of a Google Meet or Microsoft Teams meeting is worked out without you setting it, and what comes out at the end. It is for professionals whose client calls are not always in English, or not always in one language — a developer with a client in Madrid, a consultant whose French client switches to English for the numbers. The calendar side is covered under every calendar, one notetaker; every language recognised, and a page for each language the notes can be written in, is on the languages page.

Do I have to set the meeting language?

No. Nothing has to be set for the language spoken — not per meeting, not per client, not per account. Each meeting's language is worked out from the meeting itself. The transcript comes out in English whatever was spoken; the notes come out in the language you chose for them.

You can tell TrackHeros which languages a client speaks, and that is optional. It narrows what the checks below expect to hear; it never replaces them. If you do list languages, list all of them, because a language left off the list is treated as less likely. A list set before the meeting is still a guess made by someone not yet in it: one client can hold a bilingual standup and an English-only review in the same week, and the meeting itself is the only evidence that is never out of date. Why so many tools ask anyway, and what a wrong answer does to a transcript, is covered in do you have to set the meeting language?

How does it work out the language?

In four stages, cheapest first. Every stage can send a meeting to the audio pass over the recording; none can keep it on the captions alone without independent confirmation.

Stage	What it looks at	What it can decide
1. Series memory	The language this recurring meeting has been held in before	An English-only series can keep its captions; a series that has been heard in a second language goes to the audio pass
2. Caption check	Whether the platform captions read as coherent text	Captions that do not read as language — the usual sign of the wrong one — send the meeting to the audio pass
3. Audio sample	A short sample of the audio, asked only "which languages are spoken here?"	Anything other than English alone sends the meeting to the audio pass
4. Probe, then a full pass	A stretch of the recording transcribed and compared with the captions; then all of it	The captions stand only if the probe agrees with them; any doubt runs the full pass

Why stages? The audio pass is the expensive part: a second transcription of the whole recording, after the call, by a model that is not tied to one language. An English meeting with clean captions gains nothing from it but the cost. So the cheap checks run first, and the expensive one runs whenever they cannot rule it out. Every doubt resolves toward the audio pass. Being wrong in that direction costs money; being wrong in the other costs the meeting.

What happens when the meeting switches language mid-sentence?

It goes to the audio pass, and the audio pass is what reads it. Live captions on Google Meet and Microsoft Teams are set to one spoken language per meeting, so a sentence that starts in Hindi and ends in English is forced into one language and comes back fluent and wrong. Captions pick one language by design.

The audio pass transcribes the recording with a speech model that detects the spoken language itself as it goes, so a switch inside a sentence is something it expects rather than noise it explains away. The same holds for French and English, Spanish and English, German and English or Arabic and English. The captions keep one job: who spoke. More in what happens when a meeting switches language mid-sentence.

Which language are the transcript and notes in?

The transcript is in English. The notes are in the language you choose.

English is the default: the summary, decisions, action items and minutes are written in English unless you choose Arabic, French, Spanish, German, Italian, Portuguese, Ukrainian or Indonesian. Set it for your whole account, for one client, for a recurring meeting or for a single call, and the most specific setting wins; the meeting page says which one decided. You can also let each meeting decide. A call held mostly in one of those languages then gets its notes in it, and a call with no clear main language falls back to your next setting rather than to a guess. Changed your mind after the meeting? The notes can be rewritten in another language. Whichever language they are in, the model writing them is told which language the meeting was held in, so it reads the transcript as a translation rather than as the speakers' own words.

The exports follow the notes. PDF, Google Doc and Notion exports and "Copy for email" use the notes language for their headings and dates, and Arabic exports read right to left. On the free plan the notes are written in English.

A meeting held in another language is transcribed in that language and then rendered into English line by line. Each line is translated on its own and keeps its speaker and timestamp, so the English transcript still says who said what and when. A line that fails to translate keeps its original words rather than disappearing. The original-language text is stored with the transcript but not shown. How that plays out for a client who reads English but speaks something else is in English notes from a meeting in another language.

Which languages does it handle?

Close to 100 spoken languages are recognised, transcribed and rendered into English; the full list is on the languages page. A language mixed with English, including inside a sentence, takes the same audio pass. The notes can be written in nine of them, and each of the eight besides English has its own page: Arabic, French, German, Indonesian, Italian, Portuguese, Spanish and Ukrainian.

No accuracy figure is published for any language, because speech recognition is not equally strong in all of them and a single number would hide that. A meeting that mixes two languages, neither of them English, gets the same audio pass, but it is not a combination TrackHeros has tested, so no claim is made about it.

The work is split by language. Meetings in more than one language, or in a language other than English, are transcribed by a Whisper model run by DeepInfra, Together AI or Groq. English-only meetings use Google's Gemini model on Vertex AI. Both go through OpenRouter, which sends a request only to a provider that keeps no copy of the content and refuses to train on it; the promises themselves are on the privacy section of the homepage. For meetings where nobody shares a first language and English is the compromise, see meetings without a shared first language.

What do I see in the app?

Very little, on purpose.

• On the transcript page: one line, "Translated to English from a conversation held in French" — or "in Spanish and English" — with the language names in your interface language.
• Notes language: one setting in your account, on each client and on each meeting, and on the meeting page a line saying which of them decided.
• Languages spoken: an optional list on your account and on each client. Leave it empty and every meeting is listened to without assumptions.
• Speaker names: the real names from the meeting platform. The captions say who; the audio says what. A line that cannot be attributed is labelled "Unknown", never guessed.

Nothing has to be set for the language spoken, and there is no badge.

What if it gets the language wrong?

The summary says so rather than inventing content. If a transcript is unusable — garbled audio, the wrong language — the meeting page leads with "This transcript could not be read reliably" and tells you to treat what follows as unverified and check the recording.

You can switch a meeting between the Project meeting and Interview profiles, or change its notes language, and either writes the notes again from the same transcript. There is no button to transcribe a meeting again, so the recording, played inside the app, is the reference. More in what to do when a notetaker detects the wrong language.

Do recurring meetings remember their language?

Yes, and they are checked again. Once a recurring series has been heard in English only, later occurrences can skip the audio pass and keep the captions. Every tenth occurrence is checked afresh rather than taken from memory.

One sighting of a second language — a single Spanish sentence in an otherwise English weekly call — sends the series back to the audio pass at once, whatever the run of clean weeks before it. It stays there until a full pass hears English alone again. A series held in another language, or in two, gets the audio pass every time. The memory belongs to one series in one client, so "Weekly sync" under two clients is two memories that never inform each other. The reasoning is in how recurring meetings remember their language.