How accurate are AI subtitles? Accurate enough to publish after a short edit, and not accurate enough to publish without one. The "99% accurate" on a subtitle generator's page is a single number measured on that vendor's test audio, under conditions the page does not describe. On your clip the result depends on audio quality, names, accent, overlapping speakers and language mixing, and the only figure that matters is the one you measure yourself, which takes about five minutes.
This guide explains what the number means (word error rate), why one number cannot describe a tool, what the main generators actually claim as of 18 October 2026, how to measure your own clip, and what to check before you publish.
What "99% accurate" is measuring
Speech recognition accuracy is usually reported as word error rate, or WER. Per the Wikipedia entry on word error rate, read on 18 October 2026, WER is the number of substitutions, deletions and insertions divided by the number of words in the correct transcript: WER = (S + D + I) / N. "99% accurate" means a WER of 1%, or one wrong word in a hundred.
Three things about WER are worth knowing before you trust a figure.
- Every error counts the same. The metric does not weight errors by how much they matter. Writing "a" for "the" costs the same as writing "fifty" for "fifteen" in a price, or misspelling the host's name. The Wikipedia article notes that "no account is taken of the effect that different types of error may have".
- It can exceed 100%. Because insertions are counted against the reference length, a very noisy clip can score a WER above 1.0. Accuracy, as 1 minus WER, can then be negative. A vendor quoting "accuracy" is quoting the good end of the range.
- It says nothing about why. WER cannot tell a substitution from a deletion plus an insertion, and it tells you nothing about whether the errors were names, numbers or filler words. Two tools with the same WER can produce drafts that take very different times to fix.
So a WER figure is a real measurement, but it is a measurement of one test set. Move the audio and the number moves with it.
Why one number cannot describe a tool
Five variables change the result more than the choice of engine does.
Audio quality. Distance from the microphone, room echo, traffic, a fan, music under the voice. YouTube's help page on automatic captions, read on 18 October 2026, says captions are generated by machine learning so "quality may vary", and lists poor sound quality and background noise among the reasons captions go wrong or do not appear at all.
Names and numbers. A recognizer has never heard your product name, your co-host's surname or your town. It writes the closest common word. Numbers slip on vowels: fifteen and fifty, thirteen and thirty. A clip that is dense with names will score worse than a clip of general conversation on the same engine.
Accent. Engines are trained mostly on a few varieties of English, and the error rate rises as speech moves away from them. The effect is well documented in public studies, and no vendor's "99%" page states which accents the figure was measured on.
Overlapping speech. Two people talking at once, laughter over a sentence, or a host interrupting a guest. The engine transcribes one stream and drops or garbles the other. YouTube lists overlapping speakers as a reason automatic captions fail.
Code-switching. A Hindi word in an English sentence, or an English brand in a Tamil sentence. An engine set to one language has no entry for the other, so it substitutes. Hinglish exists as a separate spoken-language option for exactly this reason; the English subtitle generator handles a stray word or two, while a clip that mixes in most sentences belongs under Hinglish.
Put those together and a tool that scores 99% on clean studio speech can score far lower on a two-person podcast recorded on a phone in a cafe. Neither number is wrong. Neither describes your clip.
What the subtitle generators claim
These are the vendors' own words on their pages, read on 18 October 2026. We quote them as claims; we have not tested them.
| Tool | Accuracy wording on the page | Where it appears |
|---|---|---|
| VEED | 99.9% Accurate | Page title of the auto subtitle generator |
| Kapwing | 99% Accurate (Free) | Page title of the subtitles page |
| Maestra | 99% Accurate AI Captions | Page title of the subtitle generator |
| HappyScribe | AI up to 85-95%, human 99% | Main subtitle generator page |
| Choppity | 95%+ accuracy | Free video caption generator page |
| CapsAI | No figure published | Every draft is edited before export |
Two observations. HappyScribe is the one vendor in the list that separates its AI figure from its human figure, and its AI range of 85 to 95% is a more realistic description of what automatic drafts do across varied audio. And none of the pages says what the figure was measured on: language, accent, audio source or test size. That is not a criticism of any one company; it is a reason not to choose between them on that number.
CapsAI does not publish an accuracy figure. The product position is simpler: the recognizer produces a draft, every word, timing and style is edited in the browser, and the export is an MP4 with the captions burned in. The 3 free minutes on a new account exist so that you can measure the draft on your own clip before you pay for anything.
How to measure your own clip in five minutes
You do not need software. You need a one-minute clip, a transcript you trust and a bit of counting.
- Pick a representative minute. Typical speaker, typical room, at least two names and a few numbers. Do not pick your cleanest clip; pick the one you would actually publish.
- Write the reference. Listen and type exactly what was said, including the names spelled your way. About 150 words for a minute of speech. This is N.
- Run the clip through the tool. Upload to the auto subtitle generator, pick the spoken language, wait for the draft.
- Count the differences. Go line by line. A wrong word is a substitution. A missing word is a deletion. An extra word is an insertion. Ignore punctuation and capitalization for this count.
- Divide. Errors divided by N. 6 errors in 150 words is 4%, which a vendor would call 96% accurate.
- Look at what the errors were. This is the number that matters for your time. If 5 of the 6 were names and numbers, the fix is a two-minute search in the editor. If they were common words scattered through the clip, the audio is the problem and the recording needs to change.
Run the same minute through a second tool if you are comparing. The difference between two tools on your own voice is a real comparison; the difference between two home-page figures is not.
What to check before you publish
Whatever the WER, these are the errors viewers notice, in the order to check them.
- Names of people, places and brands. Fix every occurrence, not just the first.
- Numbers. Prices, dates, times, percentages, phone numbers.
- Negations. A dropped "not" or "never" reverses the sentence. Search the draft for lines that read oddly confident.
- Sound-alike pairs. Their and there, won't and want, and the pairs your accent produces.
- Repeated words and false starts. Delete the repeat unless it is part of the delivery.
- Line breaks. A break in the middle of a phrase slows reading even when every word is right. When to split subtitles into two lines covers the rules.
- Timing. A correct caption that arrives a second late reads as wrong. Scrub a few lines and confirm they appear with the words.
On a 60-second clip this pass takes three to five minutes in the editor. On a ten-minute talk it takes twenty to thirty. That time is the real cost of AI subtitles, and it is smaller than the time to type the transcript from scratch, which is the honest answer to "what are the benefits compared to manual captioning": you edit instead of typing, and the timing comes for free.
Accuracy in Indian languages
The same logic applies to Hindi, Tamil, Telugu, Bengali, Marathi, Gujarati, Kannada, Punjabi and Urdu, with one extra variable: whether the tool supports the language as a spoken language at all, rather than translating from English. Indian language subtitle support compared lists which generators handle which languages natively. A tool that offers Hindi as a translation target from an English transcript inherits every error in the English draft and adds its own.
What to do next
Measure before you choose. Upload a representative minute to the auto subtitle generator, count the errors against your own transcript, and look at what kind they were. A new account has 3 free minutes to start, which is three test clips. If the fixes are names and numbers, the editor handles them before export, and one-time packs start from $4 (₹49 in India, €4 in the euro area), with watermark-free export on every paid pack.
