You filmed in one language and your audience reads another. That is the whole problem, and it has a shape most tools get wrong: they solve it by sending your audio somewhere.

What this actually does

You speak. The phone writes English.

Not "captions in your language, then translated afterwards" as a second pass over a finished transcript. The speech model that hears the audio is the thing producing English, in one step, on the device. That matters for accuracy on the languages where a word-for-word transcript would be the weak link.

A video frame with captions burned in, one word highlighted as it is spoken.

It is a labelled switch, not a mode that turns itself on. That is deliberate: captions that quietly change language are worse than no captions, because you do not find out until someone tells you.

How to do it

  1. Open the clip in Zella on iPhone and go to Studio.
  2. Set the spoken language to the one you actually spoke. Do not leave it on auto if you know the answer.
  3. Turn on English captions (translate), directly under the language picker.
  4. Tap Generate. The captions arrive in English, styled with whichever of the 9 presets you are using.
  5. Edit them like any other caption track: fix a word, restyle, reposition.

Where the audio goes, which is the part worth reading

Nowhere. The model runs on your phone, there is no account, and the feature works with the network off.

That is not the norm. CapCut's own privacy policy describes pre-uploading your content before you post it: "When you create User Content, we may upload or import it to the Services before you save or post the User Content." The same policy says that if you use auto captions, "we will collect the transcribed text and the edited text," and that it may collect "the text of the words spoken in your User Content."

Read that last one twice. It is not only the video. It is a record of what you said.

For a lot of work that is merely untidy. For an unreleased product demo, a client's footage, a lesson you have not published, or anything covered by an agreement you signed, it is the difference between a tool you can use and one you cannot.

When to use this instead of captioning in your own language

Zella transcribes 27 spoken languages and captions in that language's own script. 18 of them are verified end to end, measured on native-voice audio rather than assumed from a model card. If yours is one of those and your audience reads it, caption in your language: it is the better result.

Use English captions (translate) when:

  • Your audience reads English even though you do not speak it in the video. This is most of the reach argument for creators outside the anglosphere.
  • Your language is one the transcriber cannot hold. Bengali and Telugu are the honest examples. We measured both properly and neither works on-device: every phone-capable model loops instead of transcribing. Zella does not claim them, and says so when you pick one. Translating the same audio to English does work, and was measured on the same take.
  • You are posting to a mixed feed where one caption track has to serve everyone.

What it costs

Nothing, on any tier. There is no per-minute meter, no credit balance and no upgrade prompt attached to it, because the work happens on hardware you already paid for. Zella Free covers it, the same way it covers the full editor and every AI feature.

The one honest limit is the free export ceiling, which applies to everything and has nothing to do with captions: 2 minutes at up to 1080p, watermark-free.

Common mistakes

  • Leaving the language on auto when you know it. Auto-detect is good, but naming the language removes the one place this can go wrong on a short or noisy clip.
  • Expecting it on the Mac. This is an iPhone feature. The Mac app captions on-device too, but the translate switch is not there yet.
  • Assuming the first pass is final. It is an editable caption track. Names, jargon and product words are exactly what you should skim and fix.
  • Filming somewhere loud and blaming the model. Everything downstream of the microphone is easier when the take is clean. Our notes on recording better video on iPhone cover the audio half.

Frequently asked questions

Does the video get uploaded to translate it? No. The speech model runs on the device, with no account and no network path. You can turn on airplane mode and it still works.

Which languages can I speak? 27 are offered, and 18 are verified end to end. The verified set covers English, Spanish, French, German, Italian, Dutch, Polish, Hindi, Tamil, Russian, Indonesian, Japanese, Korean, Chinese, Thai, Turkish, Vietnamese and Arabic.

Can I get captions in my own language instead of English? Yes, for the languages above, and that is usually the better choice when your audience reads them. Bengali and Telugu are the exceptions: they are measured failures on-device, and English captions are the working path for both.

Is this the same as translating captions into other languages? No, and the difference is worth keeping straight. That feature takes a finished caption track and rewrites it into 19 target languages using Apple Translation. This one changes what the speech model outputs in the first place.

Does it cost anything? No. It is in the free tier with no meter attached.

Is it available on Mac? Not yet. On iPhone only.