How well it does Urdu
Urdu needs the biggest model. With Best (Whisper large-v3 turbo, picked automatically on computers with WebGPU) we measured about 8% character error on real Urdu read speech: complete, readable Urdu where roughly one word in four has a spelling slip (خفیب for خفیف, موں for منہ). Plan a read-through before you publish. Turbo also recognised Urdu by itself, where the smaller models mistook it for Hindi. Without WebGPU the tool falls back to Accurate, which makes more mistakes.
| Model | Character error | Word error | Speed (WebGPU) | Notes |
|---|---|---|---|---|
| Best (Whisper large-v3 turbo) | 8% | 24% | 0.65× real time | Readable; fix spellings before publishing |
| Accurate (Whisper small) | 10% | 34% | 0.68× real time | About one word in three misspelt |
| Fast (Whisper base) | 16% | 36% | 0.42× real time | Not recommended for Urdu |
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 Urdu captions
- Use a computer with Chrome or Edge so the tool can run Best. On a phone it runs Accurate; the editor will say which model it used.
- Pick Urdu by hand rather than automatic if you are not using Best: the smaller models can mistake Urdu for Hindi and write it in Devanagari or Latin letters. The editor warns you when the script looks wrong.
- Keep Right-to-left fix for players ticked when you download, so players keep
۔and؟on the correct side, even on rows that start with an English word or a number. - The editor shows Urdu in a Nastaliq font where your system has one. The ASS file names Noto Nastaliq Urdu: install it (free, SIL OFL) for the same look in VLC or mpv.
- English words said in Urdu speech (common in vlogs) usually come out in Urdu script. Change them by hand if you want them in Latin letters.
Questions
Does it write Urdu script or Roman Urdu?
Urdu script (Nastaliq-style Arabic letters). Roman Urdu isn't supported; Whisper writes what it hears in the language's own script.
How accurate is it for Urdu?
With Best, about 8% of characters were wrong in our test on read speech; most errors are spellings an Urdu reader fixes in seconds. Casual speech, music under the voice and Punjabi or English mixed in make it worse.
Will it confuse Urdu with Hindi?
Best recognised Urdu correctly in our tests. The smaller models sometimes heard it as Hindi; picking Urdu by hand fixes that.
Why does the first run take a while?
The Best model is 564 MB and downloads once. The progress bar shows the size, speed and time left; after that your browser keeps it.
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.
Can I get just the Urdu text?
Yes. Switch to Transcript for paragraphs with optional timestamps, copy all, or download TXT or Word (DOCX). See video to text.