How to caption a video that mixes two languages

2026-07-25 · 7 min read

If you speak two languages in one video — English and Spanish, Hindi and English, Nepali and English — you have probably watched an auto-caption tool confidently produce nonsense. Not a blank, not a question mark: real English words that were never said. This is the single most common complaint about automatic captions, and it is not a bug you can fix by re-running it.

Why auto-captions break on code-switching

Almost every captioning tool asks for the language of your video — one choice, one file. That choice sets the model used for every second of audio. When you switch languages mid-sentence, the audio does not stop matching the model; it just starts matching it badly. The system still has to output its best guess in the language you picked, so it returns the nearest-sounding words in the wrong language.

That is why the failure looks so strange. A missing word would be obvious. Plausible English words in place of a Nepali phrase are worse, because they read as though the speaker said something they did not. Some captioning vendors document this openly and advise against using automatic captions on mixed-language content at all.

The rule: caption what was said, in the language it was said

Accessibility guidance is consistent on this. If your audience is bilingual, each portion of speech should be captioned in its own language — not translated into the majority language. Code-switching carries meaning. It is part of tone, humour, and who the person is talking to. Translating it away produces a caption that is technically readable and personally wrong.

What people do today (and why it hurts)

The workaround that circulates on support forums is genuinely this laborious:

• ✦Auto-caption the video in the majority language and export the subtitle file.
• ✦Read through it and find every passage where the speaker switched languages.
• ✦Retype those passages by hand in the correct language.
• ✦Mark them — italics, or the ISO language code in brackets — so viewers know.
• ✦Re-upload the corrected file and check the timings still line up.

For one short video that is an irritating evening. For anyone posting weekly in two languages it is not a workflow at all — it is the reason those creators either post uncaptioned or stop mixing languages on camera, which is a real cost to how they sound.

Three ways to show a non-English line

Once the speech is transcribed correctly, you still have a presentation choice, and it matters more than people expect:

• ✦Romanized — the language written in Latin letters ("maile bhane"). Usually the right default for social video: bilingual viewers read it fast, and it keeps the speaker's voice intact.
• ✦Native script — Devanagari, Chinese characters, and so on. Best when your audience reads that script comfortably.
• ✦Translated — full English. The version for people who do not speak the language at all.

There is a technical wrinkle worth knowing: translated captions cannot have word-by-word karaoke highlighting. A translation is a new sentence, so its words have no timestamps in the original audio. Any tool showing per-word highlighting on translated text is guessing. Romanized and native-script captions keep the real timings, so they can highlight honestly.

How TailorClip handles it

TailorClip detects language per phrase instead of once per file, across 90+ languages, so a sentence that starts in English and finishes in Spanish is captioned correctly on both sides of the switch. Romanized is the default; native script and translated are one tap on a finished edit, with no re-upload and no re-edit.

Two things that matter in practice. Proper nouns — your brand, your town, a friend's name — are what any transcription gets wrong most often, so you can fix a word in the transcript and the captions follow, and it learns your spelling for next time. And because the three caption modes are re-derived from the same stored transcript, you can ship one video to a home audience in romanized letters and the same cut translated for everyone else, having edited it only once.

Common mistakes

• ✦Picking the majority language and hoping. This is the one that produces invented words. If the tool only offers one language for the file, it cannot caption a switch.
• ✦Translating everything into English. Readable, but it erases the code-switching — and bilingual viewers notice immediately.
• ✦Trusting per-word highlighting on translated text. Those timings are not real.
• ✦Leaving proper nouns as transcribed. Names and brands are the highest-value five seconds of correction you will make.