How to transcribe a meeting where people switch languages mid-sentence

Most meeting platforms ask you to name the language of the meeting before it starts, then transcribe the next hour as if that answer stayed true. In a call where a German sentence carries three English nouns, or where two colleagues answer in English while the rest of the room continues in German, it does not. What you need is transcription that identifies the language segment by segment rather than once at the beginning. The rest of the work is knowing which hints help, and which failures to look for afterwards.

What switching languages does to a transcript

Linguists call it code-switching, and speech recognition research treats it as its own hard problem. The 2024 paper Aligning Speech to Languages to Enhance Code-switching Speech Recognition puts the core issue plainly: switching languages within a speech signal "results in language confusion for automatic speech recognition". The model has to decide, continuously, which language's vocabulary the next sound belongs to, and both answers are plausible for a while.

It shows up in three recognisable shapes.

One turn becomes two. A speaker starts a sentence in German, drops into English for a clause, and comes back. If the system resolves the language at the boundary rather than within the sentence, the single turn is cut into separate blocks, sometimes with a repeated fragment where the two decisions overlap.

Terminology reverts. A term the speaker used in English gets rendered with a German spelling, or a German compound is written as if it were an English phrase. This is the most expensive failure in practice, because it lands on exactly the words that carry the meeting: product names, feature names, legal and clinical terms, acronyms.

Speaker labels break at the switch point. Automatic speaker separation groups voices by how they sound, not by who they are, and people change pitch, pace and register when they change language. So the sentence where the language changed is the likeliest place in a bilingual transcript to find the wrong speaker attached to it. Worth knowing before you hunt for a labelling bug that is really a language switch.

Newer approaches attach a language label to individual output tokens rather than to the whole recording, as in the Unified Model for Code-Switching Speech Recognition and Language Identification (CALCS workshop, 2023). That is worth knowing when comparing tools: a transcript that can show you which language each part was spoken in comes from a system that decided per segment, not once.

What each platform actually allows

Sources, all accessed 22 September 2026: Microsoft, multilingual speech recognition in Teams; Zoom, viewing captions in another language, which states that "by default, English is set as the speaking language the captions are generated in"; Google, translated captions in Meet; Otter, supported transcription languages, whose help centre states that it transcribes one language at a time and that you pick it before the recording.

Teams gets closest of the four: a German speaker and a French speaker really can sit in the same meeting and each be transcribed in their own language. But it is one language per person, fixed for the meeting, and behind a paid licence. The colleague who answers a German question in English, because the term only exists in English, is still described by a setting that says German. That is the everyday case in a DACH company, and none of the four platform models covers it.

Why "one language per speaker" is not the same as multilingual

The two models differ in where the decision is made. A per-speaker model asks a person to declare a language and holds them to it. A per-segment model asks the audio, over and over, and may change its mind halfway through a sentence.

Kalima uses the second. You name the one to three languages that may come up, and Kalima decides segment by segment which of them it is hearing. Speech is recognised in 60 languages, the same list for live sessions, uploaded files and translation targets, and when the spoken language changes mid-conversation the transcript carries a small language badge on the affected segment. Finding the places where the conversation switched is then a matter of looking down the badges.

The recording itself runs on your own device, from the microphone, the computer audio, or both mixed together. Nothing joins the call and no extra name appears in the participant list, which is why it works the same way in Teams, Zoom, Meet, Webex, a Slack huddle or a room with one laptop on the table. For a distributed team whose meetings drift between languages depending on who is in the room, that is the practical difference: there is nothing to reconfigure per meeting.

Recording a mixed-language meeting so the switches survive

List only the languages that are really there. With the padlock closed, recognition sticks to the languages you picked, three at most. That restriction is what makes the transcript accurate, and it is also why a language nobody speaks costs you: it is one more wrong candidate at every ambiguous moment. Two accurate languages beat three hopeful ones.

Use session context for the terminology, not the language list. The languages answer "which language", context answers "which words". A product codename, a surname, an acronym that only means something on your team: those belong in session context, and they are the words most likely to be mangled at a language boundary.

Detection, badges, the language padlock and translation are on every plan, including the free one, so a real bilingual meeting is a cheap test. The plan differences are recording time, session length and AI usage.

When the room needs to read along

If the point is that everyone follows the meeting live, detection alone is not enough, because half the room still cannot read half the transcript. That is what live translation is for, and the mode matters. One-way renders everything into a single language, which suits a presenter and an audience. Two-way handles two languages in both directions at once and shows them side by side, so each side reads the other's words in their own language. For a German and English meeting where both sides speak, two-way is the right setting.

For the two-screens case: turn on sharing before you start, and the share link puts the same live view on a second screen in the room, or on each attendee's own phone, with nothing to install and no account required. Live viewers never consume a seat, so one device recording covers an all-hands, a community event or a talk without any AV hardware. The full treatment is in Live translation for multilingual meetings.

Tell people you are transcribing

Say at the start of the meeting that you are transcribing, in both languages if the room is mixed. It costs one sentence and removes the question later. In Germany, § 201 StGB protects the spoken word and recording a conversation without the consent of those taking part is a criminal offence, so consent from everyone present is the safe path. Check your employer's policy and the platform's terms too, and if the meeting touches personal data or works-council territory, take it to your data protection officer rather than a blog post.

Frequently asked questions

Why did my Teams transcript come out in the wrong language?

Teams transcribes according to a spoken language setting, not according to what it hears. If that setting says English and the meeting ran in German, the transcript is the engine trying to hear German words as English ones. Fixing the setting helps the next meeting, not the transcript you already have. Multilingual speech recognition, which lets each attendee declare a different spoken language, needs a Microsoft Copilot licence or Teams Premium and is not available in town halls.

Can one recording contain two languages?

It depends on the tool. Otter's help centre states that it transcribes one language at a time, chosen beforehand, and Zoom and Google Meet work from a single speaking or meeting language. Kalima detects the language per segment, so one session can hold several languages, each marked with a badge where it occurs.

Should I add every language my team speaks?

No, and with the padlock closed you can pick three at most. Add only the languages that will actually be spoken in that session. Every absent language is a wrong candidate at exactly the ambiguous moments, so two correct languages beat a long list.

Will the speaker labels be wrong in a bilingual meeting?

They are likelier to be wrong at the switch points than anywhere else, because speaker separation groups voices by how they sound and voices change character with the language. Check the labels around each language badge first. Renaming a speaker in the Studio applies across the session.

Does it translate the mixed parts automatically?

Only if you turn translation on. Detection gives you an accurate transcript in whatever was spoken, translation adds a second text alongside it. Two-way mode shows both languages side by side, which is usually what a bilingual meeting wants.

Can I get subtitles in both languages?

One track at a time. When a session has translations, the SRT and VTT export lets you choose whether the caption file holds the original or the translation, so exporting twice gives you both files.

Related

• Live translation: one-way and two-way translation, and what each is for.
• Live transcription: how the live transcript works, on any meeting platform.
• Languages and detection: the full language list, badges and expected-language hints.
• Live translation for meetings: setting up one-way or two-way before you record.
• Transcribe a Microsoft Teams meeting: without a Teams admin or a Premium licence.
• Kalima vs Otter.ai: language coverage and where the two differ.