How well it does Arabic
On a computer with WebGPU the tool uses Best (Whisper large-v3 turbo) for Arabic; elsewhere it uses Accurate. On real Modern Standard Arabic read speech Best wrote fluent, correctly spelt Arabic with about 4.5% character error, down from 7% with Accurate; most mistakes were a dropped ال or a number heard differently. Fast is several times worse.
| Model | Character error | Word error | Speed (WebGPU) | Notes |
|---|---|---|---|---|
| Best (Whisper large-v3 turbo) | 4.5% | 17% | 0.47× real time | Fluent; numbers are the usual slip |
| Accurate (Whisper small) | 7% | 24% | 0.46× real time | Fluent; check numbers and names |
| Fast (Whisper base) | 15% | 43% | 0.30× real time | About one word in four garbled |
Measured 2026-11-03 on real read speech from Google FLEURS (CC BY 4.0), in Chrome with WebGPU. Lower is better. Your results depend on the recording.
Tips for Arabic captions
- Keep Right-to-left fix for players ticked when you download: it adds an invisible right-to-left mark to each Arabic row so players keep
؟and،on the correct side, even on rows that start with a Latin word or a number. - The ASS file names Noto Naskh Arabic as its font. Install it (free, SIL OFL) for the same look in VLC or mpv, or the player uses its own Arabic font.
- Word error rates look high for Arabic because one missed ال or و counts as a whole wrong word; the character error is the better guide.
- We tested Modern Standard Arabic. Dialects (Egyptian, Levantine, Gulf, Maghrebi) vary more: read every line before you publish.
Questions
Do the subtitles show right to left?
Yes, in the editor, in the preview over your video and in the downloaded files. The right-to-left fix (on by default) makes players that guess direction from the first character show the line correctly.
Does it work with Arabic dialects?
Whisper understands many dialects but writes them closer to Modern Standard Arabic and makes more mistakes. We measured Modern Standard Arabic only.
Why does the first run take a while?
The speech model downloads once: 564 MB for Best, about 590 MB for Accurate. The progress bar shows the size, speed and time left; after that your browser keeps it and captions start in seconds.
Is my video uploaded?
No. It is transcribed in your browser. Only the optional hook writer sends the caption text, and only when you click it.