A Caption Picks Your Language in Its First Thirty Seconds.
Before a caption can write a word it has to decide what language you are speaking. That decision is made once — from thirty seconds of audio, or from a dropdown someone set last month — and it never revises itself.
Halfway through a sentence, the speaker changes language. The caption does not stutter, does not blank out, does not ask. It keeps producing lines, in the language it settled on earlier, spelling the second half the way that language would spell those sounds.
Nothing in the audio broke. The microphone was fine. What failed happened before the first word was written, and it happened once.
A decoder cannot start without choosing
Speech recognition is not a transcript of sounds. It is a search for the most likely sequence of words, and that search is weighted by a language model: which letter combinations are possible, which word usually follows which. Those weights are language-specific. English expects “the” before a noun; Polish does not have articles at all; Japanese does not put spaces between words.
So something has to answer “which language” before the first token comes out. There are only two places that answer can come from: a person, in a settings menu, or the engine, guessing.
Both produce a commitment. Neither produces a question.
The guess is thirty seconds long and it is final
The clearest published example is Whisper, the open model whose code anyone can read. When no language is specified, it runs a detection pass over up to the first thirty seconds of audio, takes the most probable language, writes that choice into its decoding options, and reuses it for the entire file. There is no second look. A ninety-minute recording inherits a decision made from its opening half-minute.
Two details make it worse than it sounds. The first: every thirty-second window after that carries the same language token — the windows are re-decoded, the label is not re-decided. The second: by default, the text produced in one window becomes the prompt for the next. One wrong word does not stay in its own window. It steers the following one, and the one after that.
That is why the output rarely looks broken. A wrong language still produces grammatical text, because the model is doing exactly what it was asked — decoding under the prior it was given. You do not get a blank line where the doubt was. You get a fluent sentence in the wrong language, or your own words spelled phonetically.
Three platforms, three places the decision hides
Commercial engines do not publish this step, but the shape is the same, because a decoder needs a prior before it can emit anything. What differs is where the decision is stored and who is allowed to change it.
Zoom puts it in a menu item called the speaking language. It defaults to English, any speaker can change it, and the official help is explicit that changing it “only changes the language the captions should be generated from” — it is not translation. Then comes the sentence that matters: the choice is saved and applied to future meetings. Set it to French once in March, and the April call starts in French, whether or not anyone is speaking French.
Teams detects the mismatch and then hands the decision to a person. If people are speaking something other than the meeting’s spoken language, Teams notices, and prompts the organizer, a co-organizer, or whoever started transcription to confirm the change. Automatic detection exists, but only in two situations: Interpreter is on, or multilingual speech recognition is enabled for the meeting — and it covers ten languages. Microsoft’s own guidance to bilinguals is the tell: speak one language at a time, stick consistently to a single language.
Google Meet is the one that moved. Until October 2026, speech translation required a single language pair configured for the whole meeting. That changed on 22 October, when Gemini-powered speech translation rolled out to 63 languages with automatic spoken-language detection, per listener, explicitly including meetings where several languages are spoken and speakers switch mid-sentence. The exclusions are just as explicit: not in live streams, not in recordings, not in client-side encrypted meetings, not on dial-in, not in third-party interoperable calls.
| Platform | Who decides | When it resets |
|---|---|---|
| Zoom | The speaker | Never |
| Teams | The organizer | Each meeting |
| Meet | The engine | Each sentence |
The third row is the newest and the least common. Auto-detection is a 2026 arrival, not a default, and it stops at the edge of the recording.
The guess is worst for the people who need captions most
Here is where it stops being a curiosity. Language identification is a classifier, and it does not hear accent as accent. It hears accent as evidence of a language.
A 2026 paper from USC and CMU describes the mechanism precisely: for non-native speech, the internal representation sits between the target language and the speaker’s first language, which produces systematic misclassification. The model is not confused by noise. It is answering a slightly different question than the one you asked — where this person is from, rather than what they are saying now.
An earlier study measured how bad it gets: identification accuracy falls sharply for accented speech and is inversely correlated with accent strength, and one measurement in that work found roughly three times as many errors on non-native speech as on native speech.
Read that consequence in the direction it actually runs. The heavier your accent, the worse your captions — and the heavier your accent, the more likely you are to need them. The failure is not distributed randomly across users. It lands on the people the feature exists for.
Why mixed speech cannot be patched
A bilingual sentence breaks the assumption on purpose. One language token per window means the window has to be assigned to one language, and a sentence that switches twice has already spent its allowance.
Three failure shapes follow, and they are worth recognising because they look like three different bugs:
- Phonetic respelling. The foreign word is written as the host language would spell those sounds. It looks like a misspelling, but the model spelled it correctly for the language it believed it was in.
- Accidental translation. The decoder, working under the wrong prior, picks the nearest equivalent instead of the word. The line reads smoothly and says something slightly different.
- Invention. When neither reading fits, the language model fills the gap with something plausible. This is the failure that produces confident sentences nobody said.
None of these is fixed by a better microphone, and none of them is visible in the accuracy number a vendor quotes you. They are decisions, not measurements.
What actually helps
Three moves, in order of how much control you have.
- Check the declared language before the call, not during it. Zoom remembers it across meetings, which means the setting you need to check is the one set weeks ago by someone else. Teams will offer the correction, but only to specific roles — if you are not the organizer or the transcript initiator, you cannot accept it.
- Ask the host to enable multilingual recognition where it exists. In Teams that is Interpreter or the multilingual option, and it is limited to ten languages. In Meet it is the translation control, available per participant, and it does not reach live streams or recordings.
- Use an engine that does not lock a language for the session. This is the only option that works when nobody in the call can change a setting, which is most of the time on someone else’s stream.
Where our half sits
The honest version of our position, including the part that is not flattering. Our standard mode pins one source language and feeds the previous text back as context — the same shape as everything above, with the same failure mode. That is why we built the real-time mode differently: it sends language hints rather than a locked choice, and decides token by token instead of committing a session.
That is a difference in behaviour, not a claim of perfection. A hint is still a bias, and a fast switch between two languages still costs accuracy. What changes is that the decision is revisited continuously instead of being fixed by the first half-minute and defended for the rest of the file.
Two boundaries, as always. It works inside a browser tab, so a meeting running in a desktop client is out of reach unless the invite offers a browser link. And it is machine translation: enough to follow a call, not a substitute for a professional interpreter when the words are contractual.
If you want to try ours, it’s called Tablingo — a browser extension that reads the tab you are already in and puts the original line above the translation. Why a bilingual episode’s transcript can only hold one language is here, what a caption tool actually reads is here, and the Zoom menu that looks like translation but is not is here.