Can AI Transcribe a Meeting When People Switch Between Languages?

Many meetings do not happen in one language. A product team in Beirut may move between Arabic, French and English in a single sentence. A sales call may start in Spanish and end in English. A family business may mix Hindi and English throughout. Can AI still produce a usable transcript?

Short answer: often yes, but the result depends heavily on how the speech system is configured. Some systems detect one main language for the whole recording. Others are designed to follow language changes word by word. Knowing which one you are using matters more than any marketing claim.

What code-switching means for speech recognition

Moving between languages in the same conversation is called code-switching. Researchers usually separate two kinds:

• Between sentences: one person finishes a sentence in English, and the next sentence is in French.
• Within a sentence: "We need the facture before Friday, yalla, let's close it."

A recent systematic review of end-to-end speech recognition research notes that switching inside a sentence is the hardest case for speech systems, and the one most studies focus on. The model has to recognize sounds, vocabulary and grammar from two languages at the same moment, often spoken with an accent shaped by a third.

Language detection is not the same as following every switch

Two provider features are frequently confused.

Automatic language detection identifies the dominant language in a recording and transcribes the audio in that language. Deepgram's documentation describes detect_language this way: it works on pre-recorded audio and returns one detected language per audio channel. It is excellent when each recording is mostly in one language you did not specify in advance, but it is not designed to report several languages inside one channel.

Multilingual code-switching mode is different. With Deepgram Nova-3, the language=multi setting uses a multilingual model that keeps transcribing as languages change and tags each word with a language code. It works for pre-recorded and streaming audio, and optional language hints can bias the model toward the languages you expect.

If a tool only detects one language, a meeting that is 70% English and 30% Arabic will usually be transcribed as English, and the Arabic portions may come out inaccurate, approximated or missing.

What ClearScribe does today

ClearScribe's web platform uses Deepgram Nova-3, but in different ways depending on the feature. Here is the current implementation, not everything the provider could theoretically do.

Recorded and uploaded meetings. When ClearScribe processes meeting audio, it requests automatic language detection, speaker diarization and utterances. The transcript stores one detected language for the recording. This works best when one language clearly dominates. ClearScribe does not currently label each language switch inside a recorded meeting transcript, so do not expect every switch to be recognized automatically and reliably.

Live Captions. ClearScribe Live Captions include a Multilingual option that uses Deepgram's multilingual Nova-3 mode, with shorter endpointing suited to code-switching. ClearScribe's current configuration lists ten languages for this option: English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch. If your meeting uses only one language, selecting that language directly is usually the better choice.

After the transcript. You can translate transcripts and work across languages, rename speakers, correct speaker assignments and generate summaries and action items. For the full list of supported languages, see ClearScribe global language support.

Speaker diarization in multilingual meetings

Speaker diarization answers "who spoke when", not "which language was spoken". Deepgram's diarization assigns a speaker number to each word, and ClearScribe turns those numbers into labels such as Speaker 1 and Speaker 2. Deepgram describes its diarization as language-agnostic, so in principle a person who switches languages keeps the same label.

In practice, diarization errors increase with overlapping speech, very short interjections, similar voices and distant microphones. Plan to review speaker labels and rename them after the meeting.

Where accuracy drops

Expect weaker results when:

• People switch languages several times within a short sentence.
• Names, product terms and acronyms are mixed into another language's grammar.
• A language or dialect falls outside the provider's supported list for the selected mode.
• Several people talk at once, or the room has echo and background noise.
• A recording is processed with single-language detection while large parts are in a second language.
• Strong regional dialects, such as colloquial varieties of Arabic, differ from the training data.

A transcript can look fluent and still be wrong. Speech models sometimes turn a foreign word into a similar-sounding word from the main language.

A practical workflow for multilingual meetings

1. Decide the main language. If one language dominates, plan around it and use single-language settings where they are available.
2. Use the multilingual mode deliberately. For live sessions that genuinely mix supported languages, choose the Multilingual option in Live Captions.
3. Improve the audio. Use a good microphone, reduce crosstalk and ask participants not to speak over each other.
4. Introduce speakers. A quick round of names at the start makes speaker review easier.
5. Review before sharing. Check switched passages, names and numbers, then rename speakers and correct segments.
6. Translate after review. Translating a corrected transcript gives better results than translating raw errors.

How to test before you rely on it

Run a small, honest test with your own meetings instead of trusting a demo:

1. Record ten to fifteen minutes of a realistic meeting with your usual participants, microphones and language mix.
2. Write a careful reference transcript for two or three minutes that contain frequent switches.
3. Compare the AI transcript with your reference. Count missing words, wrong words and switched passages rendered in the wrong language.
4. Check speaker labels separately from the words.
5. Repeat the test with a single-language setting and, where it applies, with the Multilingual Live Captions option.
6. Test again whenever your audio setup or language mix changes.

If errors are concentrated in names and terms, the review effort will be small. If whole passages in the second language are missing, that configuration does not suit the meeting.

Provider capability vs. ClearScribe feature

The bottom line

AI can transcribe multilingual meetings, and multilingual models are improving quickly. The key question is whether a product only detects one dominant language or actually follows the switches. Today, ClearScribe follows switches in Multilingual Live Captions for the listed languages, while recorded meetings use dominant-language detection with speaker diarization. Test with your own audio, review the transcript and choose the mode that matches how your team actually speaks.

Sources

• Deepgram Documentation: Multilingual Code Switching
• Deepgram Documentation: Language Detection
• Deepgram Documentation: Diarization
• Deepgram: What is speaker diarization?
• Agro, M. T. et al.: Code-Switching in End-to-End Automatic Speech Recognition: A Systematic Literature Review, arXiv:2507.07741