YouTube Auto-Captions Are 90% Accurate — And Still Unreadable
Key findings
- Median word error rate: 9.9% — about 90% of words are transcribed correctly.
- But 20% of auto-captions contain not a single period, question mark, or exclamation point.
- 32% contain no commas at all.
- Auto-captions have 90% fewer capital letters than human transcripts of the same videos.
- Among the most accurate transcripts (WER under 10%), 27% still had zero sentence punctuation.
Sample: 264 English videos, 997,401 words of human-written captions as ground truth.
Ask anyone why they don't just use YouTube's built-in transcript, and you'll hear the same answer: "the auto-captions are bad." We wanted to know what "bad" actually means — so we measured it.
The answer surprised us. The speech recognition is genuinely good. What's broken is everything that turns recognized words into readable text.
What we measured
When a creator uploads their own caption file, YouTube stores it alongside the machine-generated one. That gives us something rare: the same audio, transcribed twice — once by a human who knew what was said, once by a machine.
We took 264 English videos where both versions existed, treated the human captions as ground truth, and calculated the word error rate (WER) of the automatic version — the standard speech-recognition metric, counting substitutions, insertions and deletions after normalizing case and punctuation.
Then we did something the WER metric deliberately ignores: we measured the punctuation and capitalization that normalization throws away.
Finding 1: The recognition is better than its reputation
| Word error rate | Videos | Share | What it means in practice |
|---|---|---|---|
| Under 5% | 78 | 30% | Essentially clean |
| 5–10% | 57 | 22% | Occasional wrong word |
| 10–20% | 55 | 21% | Usable, needs a pass |
| 20–35% | 52 | 20% | Meaning survives, details don't |
| Over 35% | 22 | 8% | Not usable |
Median WER came in at 9.9%, mean at 14.8% — the gap tells you the distribution has a long bad tail rather than a uniformly mediocre middle. Half of all videos land under 10%, which for most reading purposes is fine. But the 10th-to-90th percentile spread runs from 1.7% to 31.5%, so "how accurate is it" has no single honest answer. It depends almost entirely on the audio.
The worst performers in our sample shared predictable traits: heavy background music, overlapping speakers, strong accents, and dense technical vocabulary. The best were single-speaker, studio-recorded, conversational English.
Finding 2: The formatting collapses almost everywhere
Here is the same set of videos, comparing structure rather than words:
| Measure | Auto-captions | Human captions | Difference |
|---|---|---|---|
| Sentence-ending marks per 1,000 characters | 8.13 | 12.86 | −37% |
| Commas per 1,000 characters | 5.90 | 8.93 | −34% |
| Capital letters (share of all letters) | 2.57% | 26.08% | −90% |
| Digits (share of all characters) | 0.51% | 0.53% | −4% |
The averages understate it, because they're dragged up by the minority of auto-captions that do carry punctuation. Counted as all-or-nothing:
| Auto-captions that contain… | Videos | Share |
|---|---|---|
| No sentence-ending punctuation at all | 54 / 264 | 20% |
| No commas at all | 85 / 264 | 32% |
| Almost no capital letters (under 1%) | 49 / 264 | 19% |
| Neither punctuation nor capitals | 23 / 264 | 9% |
One in five transcripts is a single unbroken stream of words with no sentence boundaries anywhere in it. Not "poorly punctuated" — unpunctuated.
Finding 3: Accuracy and readability are not the same problem
This is the result that reframes everything. We took only the best transcripts — the 135 videos with word error rates under 10%, where recognition was demonstrably working — and checked their formatting.
27% of them had zero sentence punctuation.
Getting the words right and producing readable text are independent achievements, and YouTube only does the first one. Here's what that looks like in practice — the same passage, machine and human:
Every word matches. The WER of that passage is zero. And it is still meaningfully worse text: you cannot skim it, you cannot quote it, and if you paste it into a language model, the model has to spend effort reconstructing sentence boundaries before it can do anything useful with the content.
Why this happens
Speech recognition systems output a token stream, not a document. The acoustic model hears sounds and maps them to words; nothing in that pipeline is responsible for deciding where a sentence ends. Punctuation and capitalization are a separate post-processing task, usually handled by a different model entirely.
YouTube's captions are also built for a specific job that isn't reading: displaying two lines at a time, synchronized to audio, for viewers who are watching the video. In that context, a period earns you nothing. The design is reasonable. It just produces something that fails badly the moment you take the text out of the player.
There's a second consequence worth noting: because caption files are cut into display-sized chunks rather than sentences, a single sentence is routinely split across three or four timed segments. Copy the transcript out and you inherit that fragmentation.
Who this actually hurts
- Students and researchers quoting from lectures — an unpunctuated block can't be cited without manual repair.
- Anyone feeding transcripts to an AI model — missing sentence boundaries measurably degrade summarization and question-answering quality, and you pay tokens for text the model has to fix first.
- Deaf and hard-of-hearing viewers — this is the group with no alternative. Auto-captions with no punctuation and no speaker labels make multi-person content genuinely difficult to follow.
- Non-native speakers — sentence boundaries are a major comprehension aid when you're reading in a second language.
How to fix it
The good news is that this is a solved problem in the literature, just not one YouTube applies. Restoring readable structure to a raw transcript involves three independent steps:
- Punctuation restoration — inserting sentence boundaries and commas based on syntax and prosody. This does not change any words.
- Truecasing — restoring capitalization for sentence starts and proper nouns. Note that this is the one place where a model can introduce errors that weren't in the original, since it has to guess whether "apple" is a fruit or a company.
- Paragraph segmentation — grouping sentences into topical blocks, which is what makes a long transcript skimmable.
None of these require re-running speech recognition. They operate on the text you already have, which is why they're fast and cheap compared to transcription itself.
Get a clean transcript instead of a raw one.
Paste any YouTube URL to extract the full transcript with punctuation, capitalization and paragraphs restored — then copy it, download it as TXT or SRT, or translate it.
Methodology, and what this study cannot tell you
We want this data to be citable, which means being explicit about its limits.
What we did
- Identified English videos where both an automatic and a human-uploaded caption track existed for the same language.
- Randomized the sample order, then took 306 pairs. After parsing, 296 were usable.
- Excluded 32 pairs where the two transcripts differed in length by more than ±60%, on the assumption that the human track was partial, abandoned, or covering different content. That left 264 videos, 997,401 words of ground truth.
- Calculated WER on normalized text (lowercased, punctuation stripped) using word-level edit distance. Transcripts longer than 4,000 words were truncated to their first 4,000 for the WER calculation (52 samples); structural measures used the full text.
Limitations
- The ground truth is not professional transcription. Human-uploaded YouTube captions vary in quality, and some are community-contributed. Real WER against a professional reference would likely be somewhat different — probably lower, since some of what we counted as machine error is human paraphrasing.
- The sample is not a random draw from YouTube. These are videos that real users of our tool requested, which skews toward content people want transcripts of — lectures, interviews, tutorials — and away from music and gaming.
- English only. Recognition quality varies substantially by language, and we would expect worse numbers for languages with less training data.
- Videos with human captions may not be representative. Creators who bother to upload captions may also produce cleaner audio, which could bias WER downward.
- WER above 100% is possible and appears in our tail — it happens when the system inserts more words than the reference contains.
Citing this study
The data is free to reference. Suggested citation:
youtube-transcript.ai (2026). YouTube Auto-Caption Accuracy and Readability Study. 264 English videos, 997,401 words of human-caption ground truth. Retrieved from https://youtube-transcript.ai/blog/youtube-auto-caption-accuracy-study
If you're writing about caption accuracy, accessibility, or AI transcription and want a cut of this data we didn't publish — by video length, by content category, or the raw WER distribution — get in touch and we'll share it.
The short version
YouTube's speech recognition is not the weak link. It gets roughly nine words out of ten right, and on clean audio it's close to perfect. The failure is that it hands you those words as an undifferentiated stream: no sentences, no capitals, no paragraphs. One in five transcripts doesn't contain a single period.
"The auto-captions are bad" turns out to mean "the auto-captions are unformatted." That's a different problem, and unlike acoustic accuracy, it's one that can be fixed entirely in post-processing.