How Accurate Are YouTube Auto-Captions? We Measured Eight Languages (2026)
How good is YouTube auto caption accuracy? On captions as served on 2026-10-26, YouTube's automatic captions had a pooled word error rate of 14.2% (95% CI 10.7% to 18.5%) across 33 captioned talks, lectures and educational videos in seven languages, measured against each creator's own caption track. Japanese, scored by character, was 13.5%. Disclosure: we sell a transcription tool. This study reports YouTube's captions only; we ran our own transcription pipeline on the same videos solely to flag outlier references, and do not report its accuracy here.
That pooled figure hides a wide spread. English was at 3.0%, while Korean was at 29.0% by words (11.6% by characters). With five videos or fewer per language, every per-language number below is indicative, not definitive. The per-language CSV, the per-video results with drop counts, and the list of scored videos are published below, so anyone can fetch the same tracks and check the numbers.
How accurate are YouTube auto-captions in each language?
The answer depends heavily on the language: on captions as served on 2026-10-26, YouTube's automatic captions ranged from 3.0% word error rate in English to 29.0% in Korean, against the creator-uploaded caption track of the same video.
| Language | Metric | Videos | Error rate (95% CI) | Without rule 3 |
|---|---|---|---|---|
| English | WER | 5 | 3.0% (2.4% to 3.7%) | same |
| Spanish | WER | 5 | 8.7% (4.4% to 13.0%) | same |
| French | WER | 4 | 12.9% (8.7% to 19.4%) | 16.9% (5 videos) |
| German | WER | 5 | 18.0% (13.2% to 22.6%) | same |
| Portuguese | WER | 5 | 16.8% (9.2% to 25.2%) | same |
| Japanese | CER | 5 | 13.5% (7.9% to 19.0%) | same |
| Korean | WER | 5 | 29.0% (17.1% to 44.0%) | same |
| Hindi | WER | 4 | 22.9% (2.0% to 49.2%) | 29.6% (5 videos) |
| Pooled (seven word-scored languages) | WER | 33 | 14.2% (10.7% to 18.5%) | 15.9% (35 videos) |
WER is word error rate: substitutions, deletions and insertions divided by the number of words in the reference. CER is the same count over characters, used for Japanese because it is written without spaces between words. For Korean and Hindi we also report CER, because word boundaries in both languages make WER harsh on small spacing or spelling differences: Korean CER was 11.6% (4.6% to 22.7%) and Hindi CER was 17.0% (0.9% to 37.0%), on captions as served on 2026-10-26.
The micro-average (total errors over total reference words) is the headline. The macro-average, which weights every video equally, was 15.4% for the pooled set. Both are in the CSV.
How did we measure YouTube caption accuracy?
We compared two caption tracks of the same public video: the one YouTube generates automatically and the one the creator uploaded. Creator captions are the reference, which makes this a measure of how far YouTube's automatic track is from what the creator published, not from a perfect transcript.
- Finding videos. We searched YouTube in each language for talks, lectures, news and educational videos using YouTube's own "Subtitles/CC" and medium-length (four to twenty minute) search filters, then kept public videos up to fifteen minutes long (so in practice four to fifteen minutes), with at most two videos per channel per language. A first discovery pass with plain keyword search found too few captioned videos (two qualifying Spanish videos in sixty tries) and was discarded before scoring. No cookies, no sign-in, no age-restricted or members-only videos.
- Qualifying a video. YouTube had to report the video's language as the target language; the video needed a creator-uploaded caption track in that language that was not labelled as a translation, and the text of that track had to actually be in that language (Hindi tracks, in particular, often turned out to hold an English translation); and YouTube had to offer its automatic speech-recognition track in the original language.
- Fetching. Both tracks were downloaded on 2026-10-26 with the open-source yt-dlp tool, version 2026.10.21. The automatic track was rebuilt from YouTube's word-level caption events, which avoids the duplicated rolling lines that appear in YouTube's VTT export.
- Normalizing. Every text was normalized the same way before scoring: Unicode compatibility folding, lowercasing, curly quotes folded, all punctuation removed, bracketed non-speech tags such as [Music] or (laughs) removed, a short fixed list of filler words removed from both sides (for example "um" and "uh" in English, "euh" in French, えー and あの in Japanese), and, for Hindi, zero-width joiners and nukta and chandrabindu variants folded. Numbers were not normalized: "10" against "ten" counts as an error. Two side effects apply to every track equally: all parenthesized text is removed, including real parentheticals such as acronyms, and Japanese fillers are removed wherever they appear, so a word that contains one (あの人) loses it too.
- Scoring. We computed WER (CER for Japanese) per video, then pooled. The 95% confidence intervals come from a bootstrap that resamples whole videos within each language, 2,000 resamples with a fixed seed, so the intervals reflect how much the result depends on which videos happened to be in the sample.
Which videos were dropped, and why?
Some creator caption tracks are not usable references. Rules 1 to 4 below and their thresholds were fixed in advance, written into the results file before final scoring, and never tuned afterwards. The first row, a check that the creator track's text is really in the spoken language, was added after a first screening pass had already computed error rates, when Hindi tracks scoring over 100% turned out to be English translations; it removes unusable references but it does lower the Hindi figure, so we flag it. Across 98 candidate videos that met the search and caption-track checks, the counts were:
| Reason | Videos dropped |
|---|---|
| Creator track text was not in the spoken language (a translation filed under the language's code) | 9 |
| Rule 1: creator track far too short for the video length (a summary or partial captions) | 3 |
| Rule 2: creator and automatic tracks differ in length by more than 30% | 3 |
| Rule 4: creator track is a re-upload of YouTube's own automatic track (under 2% different) | 27 |
| Rule 3: three-way outlier (see below) | 2 |
Rule 4 matters more than it looks. Twenty-seven creator tracks, nineteen of them in Hindi, were near-identical copies of YouTube's own automatic captions. Scoring those would grade YouTube against itself and push its error rate toward zero. After the drops, each language kept its first five qualifying videos in search order; Japanese and Hindi needed deeper searches than the others to reach five.
Does dropping outlier references change the result?
Only a little: the pooled word error rate is 14.2% with rule 3 applied and 15.9% without it, on captions as served on 2026-10-26. Rule 3 drops a video when YouTube's automatic track and our own transcription pipeline agree closely with each other (under 15% apart) while both are more than 40% away from the creator track. In that case the creator track is the likely outlier, for example a heavily edited or paraphrased caption, rather than both machines making the same mistakes.
That rule can only fire when YouTube's error is high, so it can only lower YouTube's measured error. That is why both figures are published. Our pipeline was used as an outlier detector only, never as a reference, and its own accuracy is not reported here. Rule 3 removed one French and one Hindi video.
What are the limitations of this study?
The main limitation is size: 38 scored videos in total, five or fewer per language, so treat the per-language rows as indicative and lean on the pooled figure and its interval.
- Captioned content only. Every video had creator-uploaded captions, which skews toward channels that invest in captioning: talks, lectures, explainers and some vlog-style educational videos. Videos nobody captions, such as casual uploads, music, gaming or noisy field recordings, were not measured and are likely harder.
- Creator captions are not verbatim transcripts. Creators sometimes tidy up their captions, removing false starts or rephrasing. Those edits count as YouTube errors here, so part of every number is editorial difference, not recognition error.
- Lightly edited re-uploads can pass. Rule 4 only catches creator tracks under 2% different from YouTube's automatic track. A creator who uploads YouTube's own captions with a few corrections can pass it. Nine of the 38 kept videos scored under 5%; some of those may be such copies, which would make YouTube look more accurate than it is.
- Numbers count as errors. "Twenty" against "20" is scored as a mistake, which raises error rates for any track that writes numbers differently from the reference.
- One point in time. YouTube updates its speech recognition. These figures describe captions as served on 2026-10-26, not YouTube's captions in general.
Where can I get the data?
The per-language results are a CSV: youtube-auto-caption-accuracy-2026-10.csv. Each row gives the language, metric, number of videos, reference word count, micro and macro error rates, the 95% interval, and the same figures without rule 3. The per-video results, thresholds and drop counts are in the results file, and the scored videos are listed by YouTube ID in the video list, so the tracks can be fetched again. Caption text itself is not republished.
When are YouTube auto-captions good enough?
For English talks with clear audio, YouTube's automatic captions were close to the creator's own track in this sample, at 3.0% word error rate on captions as served on 2026-10-26. For the other seven languages, expect to correct a noticeable share of words, and for Korean and Hindi, a large share.
If you only need the existing captions of a video, a caption downloader is enough. If the video has no good captions, or you need speaker labels and timestamps, a tool that transcribes the audio itself is the other route: our free YouTube transcript tool does that, and our guide to the YouTube transcript generators in 2026 explains the difference between scraping captions and transcribing audio.
Frequently asked questions
How accurate are YouTube's auto-generated captions?
In this study, on captions as served on 2026-10-26, YouTube's automatic captions had a pooled word error rate of 14.2% (95% CI 10.7% to 18.5%) against creator-uploaded captions across 33 talks, lectures and educational videos in seven languages. That is roughly one word in seven differing from the creator's track. Japanese, scored by character, was 13.5%.
Are YouTube auto-captions accurate in English?
Yes, for clear talks and lectures. English had a word error rate of 3.0% (95% CI 2.4% to 3.7%) across five captioned videos, on captions as served on 2026-10-26. That is about three words in a hundred differing from the creator's own captions.
Are YouTube auto-captions accurate in Korean and Hindi?
Less so. In this sample, on captions as served on 2026-10-26, Korean had a 29.0% word error rate (11.6% by characters) and Hindi 22.9% by words (17.0% by characters) against the creator tracks. Both rest on four or five videos, with wide intervals, so these figures are indicative. Hindi especially: two of its four videos sit just above the 2% re-upload line and one scored 60.8%, so the Hindi figure could move a lot with a different sample.
What is a good word error rate for captions?
There is no single standard, but as a rough guide, under 5% reads as nearly clean and over 20% needs heavy editing. Nine of the 38 videos in this study scored under 5%, most of them in English.
Did transcribe.so measure its own accuracy in this study?
We ran our own transcription pipeline on the same videos, but only to flag creator tracks that were likely outliers (rule 3). This study does not report its accuracy: it measures YouTube's automatic captions only. We sell a transcription tool, so treat that as a disclosed interest.