Unsure words go yellow
About 10% of words get flagged, and those flags catch roughly half of all errors, so a quick scan of yellow words fixes most mistakes.
Home / Accuracy
Benchmark · Oct 2026How accurate are the captions?
Short answer: on 28 real, noisy, accented English clips, the English engine in Offline Captions (NVIDIA Parakeet-TDT 0.6B v2, run on your computer) got a 7.6% word error rate, about 1 word in 13, and better than every Whisper model we tried. On real Hinglish creator speech it got 22–24%. We haven't finished a same-clip test against CapCut or Submagic, so we don't claim to beat them. Here's exactly how we measured.
English results
| Engine | Used in | Word error rate | Normalised WER | Speed (2 vCPU) |
|---|---|---|---|---|
| NVIDIA Parakeet-TDT 0.6B v2 | ✓ English default | 7.6% | 6.2% | 17.5× realtime |
| Whisper large-v3-turbo | 8.4% | 7.1% | 0.8× (too slow on CPU) | |
| Whisper small | Web version | 9.6% | 8.3% | 3.0× |
| Whisper base | Web "fast" | 14.2% | 12.9% | 6.4× |
Data: 28 clips (330 seconds) from the People's Speech test set: internet audio, noisy and accented. WER excludes "um/uh/you know". Normalised also folds reference-format quirks (gonna → going to, compounds, "percent"). Tested 16 Oct 2026.
Hinglish results
| Engine | Hinglish mode | Word error rate (script-blind) |
|---|---|---|
| Hinglish "accurate" (Apex) | ✓ | 22.3% |
| Hinglish "fast/balanced" (Swift) | ✓ | 23.6% |
| IndicConformer Hindi (native-script model) | 30.4% |
Data: 10 real creator clips from CoSHE-Eval. Script-blind means "complaint" and "कंप्लेंट" count as the same word, so the score measures hearing, not spelling. About 80% of words get exact word timing; the rest are timed by estimate and flagged.
Fixing the errors that remain
Names fixed once
Most remaining errors are proper names. Add them once; near-misses snap to your spelling and are marked for a check.
What we don't claim
No same-clip comparison with CapCut or Submagic yet. Heavy music, crosstalk and strong echo still hurt every engine, ours included.
Run it yourself
The benchmark scripts score any clips you have human transcripts for: word error rate, clips needing two fixes or fewer, how many wrong words the flags caught, and speed. Send us your results, good or bad, and we'll publish them here. A head-to-head against CapCut and Submagic on the same clips is in progress and will be added with its raw files.
FAQ
How accurate is Offline Captions?
On 28 noisy, accented English clips from the People's Speech test set, its default engine (NVIDIA Parakeet-TDT 0.6B v2) had a 7.6% word error rate (6.2% normalised). Hinglish mode scored 22–24% on real creator speech (script-blind).
Is Offline Captions more accurate than CapCut?
We don't know yet and don't claim it. A same-clip test against CapCut and Submagic is in progress; results will be published on this page.
Are offline caption models as good as cloud ones?
For English they are now close: open models like NVIDIA Parakeet run faster than realtime on a laptop CPU with single-digit error rates on real-world audio in our test.
Stop renting your captions.
Unlimited videos, offline, no watermark. $39 once, or $29 as a founding pre-order.