Can AI transcribe two languages in the same sentence?

The short answer

It depends on where the app decides the language. There are two designs. One decides once, for the whole recording, usually because you told it. The other decides continuously, word by word, from the sound itself. Only the second design can handle a sentence that changes language halfway through.

Apps that decide per recording: Notta and Otter.ai make you choose the language before you start. Microsoft Teams' own transcript works the same way, with a spoken-language setting. Whatever you choose, the other language suffers.

Apps that decide per word: Noter AI runs automatic language identification on every recording with no selector at all, and its model tags language at the word level. Fireflies also auto-detects and copes with switching, with the caveats covered below.

Code-switching is the technical name for mixing languages inside one conversation, and it is normal speech for bilingual teams, families running a business, and any meeting with an interpreter. It is not an edge case. It is how a large share of the world's meetings actually sound.

Why most apps fail at this

A speech recogniser turns sound into text using a model trained on a language. When the app asks you to choose Spanish, it loads a Spanish model and treats every sound as Spanish. An English phrase in the middle of that recording is still matched against Spanish words, so "let's ship it Friday" becomes a string of Spanish syllables that never happened.

The failure is silent. The transcript still looks like a transcript. It has sentences, punctuation, speaker labels. You only notice when you read the part you care about and it makes no sense. Then the summary built on top of it inherits the same garbage, and the action items with it.

Switching the language setting mid-meeting does not fix it. Some tools let you change the language during a live session. That helps when the meeting moves from one language to another in blocks. It does nothing for a sentence that mixes both, because the model still only holds one language at a time.

How word-level detection works

A multilingual model is trained on many languages at once and predicts, for each short slice of audio, both what was said and which language it belongs to. There is no single language chosen up front, so a change of language is just another prediction, made continuously.

Noter AI uses this design. Transcription runs on a unified speech model covering 60+ languages, with automatic language identification on for every recording. You can give it hints about which languages you expect, which nudges accuracy, but hints never lock anything. The output is one transcript with each word written in its own script, speaker labels and timestamps included.

What it still cannot do: two people talking over each other in two languages at once is as hard as any overlapping speech, and a single borrowed word inside a sentence ("the deadline is الخميس") is sometimes normalised to the surrounding language. Full phrases and clauses are where the design earns its keep.

Which apps handle mixed-language sentences

Two apps do the job. Between them, the deciding factors are usually whether the meeting is in a room and what happens to the notes afterwards. Noter AI records in person with one tap and translates the summary and action items rather than the raw transcript. Fireflies has the larger language catalogue, at 100+, but rations its summaries with monthly credits and charges per seat.

A two-minute test before you subscribe

1. 1Record yourself for one minute switching languages mid-sentence, the way your team actually talks. Include one full sentence in each language and two that mix.
2. 2Do not set a language. If the app will not let you record without choosing one, that is your answer already.
3. 3Read the mixed sentences. Both halves should be in their own script and readable. Phonetic spelling of the second language means the app locked to one.
4. 4Check the summary. Run whatever summary or action-item feature the app has. If the transcript was wrong, the summary will be confidently wrong too, which is the more expensive failure.
5. 5Try it in the room as well as on a call. Video calls give the app a clean audio stream. A phone on a table is the harder case and the one most bilingual meetings actually happen in.

What Noter AI gives you

No language selector. Automatic language identification runs on every recording across 60+ languages. Start recording, put the phone on the table, and let people talk the way they normally do.

Word-level language tagging. A sentence that begins in one language and ends in another comes back with both parts correct, in one transcript, with speaker labels and timestamps. Tap any line to hear that moment.

The summary itself is translated. Choose a target language and the summary, analysis and action items are produced in it. The original stays too. Translation covers 18 languages; transcription covers 60+.

Both halves of your week. In-person recording from an iPhone, iPad or Android phone, and a bot for Zoom, Microsoft Teams, Google Meet and Webex.

One flat price. $9.99 a month or $69.99 a year, no per-seat multiplication and no monthly credits.

The honest limits. No web or desktop app, no CRM integrations, and no permanent free tier. Overlapping speech in two languages is still hard, as it is everywhere. Run the two-minute test above with your own team before subscribing.