Caption Languages & Models: Bangla, Hindi, Chinese & More
Version note: everything on this page is Zella 1.0.7 or later, which is out now on the App Store and the Mac App Store. On 1.0.6 or earlier, captions work as described in captions & subtitles and Bangla is not supported.
Quick answer: From Zella 1.0.7, a captions model comes with the app on both Mac and iPhone, so captions work the moment you install, offline, in English, Spanish, Portuguese, French, German, Chinese, Japanese, Korean, Arabic, Russian, Indonesian and more. Bangla and Hindi (with its neighbours) write in their own script only on a bigger model, and Zella offers it free, once, the first time you caption in one of them. Say no and it captions on the included model anyway.
On this page: what comes with the app · which language uses which model · the free upgrade pop-up · Bangla · Chinese script · auto-detect · measured results · FAQ
What comes with the app
Zella 1.0.7 ships one captions model inside the app. It is called Standard. It is why captions need no download and no internet on first launch, and why nothing about your video is ever sent anywhere to be captioned.
Two bigger models are optional, free, and live on our own server, models.zellahq.com:
| Model | Size | Comes with the app? | What it is for |
|---|---|---|---|
| Standard | about 150 MB | Yes | English, Spanish, Portuguese, French, German, Chinese, Japanese, Korean, Arabic, Russian, Indonesian and most other languages |
| Best | about 465 MB (iPhone) · 466 MB (Mac) | No, free download | More accurate in most languages, and needed for Hindi, which Standard writes in the wrong script |
| Bangla | about 485 MB (iPhone) · 539 MB (Mac) | No, free download | Bangla (বাংলা), written in Bangla script |
A bigger model downloads once. After that it works offline like everything else.
Which language uses which model
You never have to choose a model by hand. Zella picks the right one for the language you set (or the one it hears, on Auto-detect):
- Bangla runs on the Bangla model when it is installed.
- Hindi runs on Best when it is installed. Urdu, Tamil, Gujarati, Kannada, Malayalam, Marathi, Punjabi, Nepali, Sinhala and Thai are sent to Best too, but we have only measured Hindi, so treat those as untested and read the captions before you export. Telugu is the exception: it was measured and does not work reliably.
- Everything else runs on Standard, or on Best if you chose Best as your model.
If the model a language needs is not on your device yet, Zella captions on Standard rather than stopping, and offers the free download.
The free upgrade pop-up
The first time you caption in Bangla or a language that needs Best, Zella shows one pop-up before it starts:
- Download free fetches the model from zellahq.com, shows the size first, and carries on captioning when it lands.
- Use the included model captions straight away on Standard, with no download.
After a caption run on Standard, Zella may offer Best once a week, with Not now and Don’t ask again buttons. It never asks while you have Best, and never asks again once you say so.
You can also pick a model yourself: Studio → Captions → Captions model on iPhone, or the Model button beside the language picker in the Captions tab on Mac.
Bangla captions
Before 1.0.7, Zella could not caption Bangla. We measured it properly: the general-purpose models either wrote Bangla in the wrong alphabet or produced repeating gibberish, and we said so in these docs rather than pretend.
What changed is a Bangla-specific model, a Whisper model fine-tuned on Bangla speech by the open-source community and converted by us to run on your device. It writes বাংলা in Bangla script, offline, and nothing leaves your device.
- Mac: a larger Bangla model, about 4.5% character error on our test sentence.
- iPhone: a smaller model (the larger one took too long to load on phone hardware), about 20% character error on the same sentence, still in Bangla script.
Both are far better than the 100% error the general models scored. Fast, clear speech works best; a busy background will cost accuracy, as it does in every language.
Credit: the Mac Bangla model is bengaliAI/tugstugi_bengaliai-asr_whisper-medium (Apache 2.0). The iPhone model is bangla-speech-processing/BanglaASR (MIT). See licences.
Chinese: Simplified or Traditional
Speech models write Chinese in whichever script they happened to favour. Zella sets it for you: Simplified characters by default, and Traditional characters when your Mac or iPhone region is Taiwan, Hong Kong or Macau.
Auto-detect
Leave the language on Auto-detect and Zella listens to the opening of your take, names the language, and then chooses the model. That means a Bangla take on Auto is captioned on the Bangla model, and a Hindi take on Best, without you picking anything.
If you pick a language yourself and the audio is clearly a different one, the Mac declines to caption and tells you, rather than inventing words in the wrong language. Close neighbours (Hindi and Urdu, Spanish and Portuguese) are never refused.
How well each language works
We measure every language on a recorded sentence read by a native-language voice, and score character error (lower is better; 0% is perfect). These were measured on 2026-10-26 on the shipping models.
| Language | Model | Character error |
|---|---|---|
| English | Standard | about 5% |
| Chinese (Simplified) | Standard | 0% |
| Portuguese | Standard | about 2% |
| Spanish | Standard | about 6% |
| French | Standard | about 4% |
| German | Standard | 0% |
| Japanese | Standard | about 10% |
| Korean | Standard | about 2% |
| Arabic | Standard | about 2% |
| Russian | Standard | about 5% |
| Indonesian | Standard | about 7% |
| Hindi | Best | about 18% |
| Bangla (iPhone) | Bangla model | about 20% |
| Bangla (Mac) | Bangla model | about 5% |
One sentence per language is a quick check, not a guarantee. Long, fast or noisy audio scores worse in every language. Telugu is still not one we can honestly promise: it was measured and fails on-device, so use English captions (translate) on iPhone for Telugu.
FAQ
Do I have to download anything to caption? No. From 1.0.7 the Standard model comes with the app. You only download if you choose to, for Bangla, Hindi and similar.
Is the download really free? Yes. Best and the Bangla model are free for every user, on every plan, from zellahq.com.
Does my video get uploaded to caption it? Never. Captioning runs on your device. The only network request is the one-time model file download, which carries nothing from your project.
I’m on Zella 1.0.6. Does any of this apply? Not yet. On 1.0.6 the captions model downloads the first time you transcribe, and Bangla is not supported. Update to 1.0.7 for everything on this page.
Can I remove a model to free up space? On iPhone, Settings → Captions model → Remove download. On Mac, delete the model file from the Zella folder in Application Support; Zella offers it again if you caption in a language that needs it.
Why does my Hindi caption look like Urdu? That is the Standard model writing Hindi in the wrong script. Install Best (the pop-up offers it) and caption again.
Related: Captions & subtitles → · Captions on iPhone → · Install & permissions → · Troubleshooting →