Bharti Subtitles

No. Nothing is sent to any API. The model runs inside your own browser.

No. The browser reads the file straight from your own Media Library and works on it locally. The only thing downloaded is the speech model itself, once.

The speech model has to be downloaded the first time — about 50 MB for most languages, about 250 MB for Hindi, Marathi and Hinglish, which need a larger one. After that the browser has it and there is no wait. Transcription speed then depends on the machine, not on your hosting — a modern laptop does roughly real time or better.

Not in this version. Everything is deliberately in the browser, which is what keeps the cost at zero and the video private. Bulk generation would mean a server doing the work, and that means both of those go away.

Check three things: the video has subtitles saved (Media > Bharti Subtitles will say so), Settings > Bharti Subtitles > "Automatically show subtitles" is on, and your theme is not replacing the video player with its own. If it is, turn the setting off and add the track yourself.

The Video block has Playback controls switched off, so the player draws no controls at all. Select the video in the editor and turn that setting on. Nothing is wrong with the subtitles — they are attached, which is why they appear on screen.

The video is in a codec your browser will not decode — usually HEVC inside a .mov. Convert it to MP4 (H.264 + AAC), re-upload, and it will work.

English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Polish, Japanese, Korean, Chinese, Turkish and Indonesian are the strong ones. Hindi, Marathi and Hinglish also work, with two honest caveats below. Only languages the on-device model actually handles are offered — listing more would just produce bad subtitles.

Two things, and the panel tells you both before you press the button. They need the larger speech model — about 250 MB instead of 50 MB, and slower to run. The smaller model does not write these languages down at all: it turns Hindi into English and Marathi into Roman letters, so the size buys the script itself. And be realistic about the accuracy. The model has seen far less of these languages than of English, and on a Marathi test most lines came back with the script correct but the words wrong — close enough to recognise what was said, not close enough to publish. Expect to rewrite most of them. What is reliably right is the timing. The tedious half of subtitling is deciding which line shows when, and that comes out correct: fourteen cues, each three or four seconds, following the speech. Retyping the words into a finished set of cues is still much faster than building the whole thing from nothing — but that is what you are getting, not a draft you lightly correct. English and the European languages do not have this problem.

The panel is for site owners, so it is built for a desktop browser. The subtitles it produces show on every device, mobile included.