How accurate are automatic captions?
By the maker of Captions Cut. Facts checked on from the pages listed under Sources.
The short answer
On clear speech in a major language, today’s best speech recognition gets about 95 to 98 words in 100 right (a word error rate of 2 to 5%). That still leaves a few mistakes in a 45-second clip, usually in names, slang and numbers. Accuracy drops sharply with background noise, distant microphones, overlapping speakers, strong accents and less-resourced languages, so read every caption before posting.
Word error rate, in plain words
Word error rate (WER) counts the words a system gets wrong (substituted, left out or added) per hundred spoken. A 5% WER means about one mistake every twenty words. A short clip has 100 to 150 words, so even an excellent system usually leaves something to fix.
What systems score
| Benchmark | Result | Who measured |
|---|---|---|
| Artificial Analysis, ~8 hours of real-world English audio | ElevenLabs Scribe v2: 2.2% WER, about fourth of the models tested | Independent benchmark |
| AssemblyAI’s own datasets | Its Universal-3.5 Pro 4.35%; GPT-4o Transcribe 5.34%; Scribe v2 5.87%; Deepgram Nova-3 6.66% | AssemblyAI (a vendor) |
| LibriSpeech test-clean (read audiobooks) | Whisper large-v2: 2.7% WER | OpenAI’s Whisper paper |
| ElevenLabs’ own language tiers | 5% or less for English, French, German, Spanish and Japanese; 20 to 50% for Urdu, Yoruba and Zulu | ElevenLabs (a vendor) |
What makes it fail
- Noise and distance. In the CHiME-6 challenge, recordings of people talking across a room scored around 51% WER for the baseline system.
- Who is speaking. A 2020 study in PNAS found five commercial systems averaged 35% WER for Black speakers against 19% for white speakers.
- Overlap and mixed languages. YouTube lists overlapping speakers, mixed languages and long silences among the reasons its own captions fail.
- Inventions. The Whisper paper notes that models can occasionally output text unrelated to the audio.
Getting cleaner captions
- Record close to the microphone, away from music and traffic.
- Choose the spoken language instead of auto-detecting it on short or mixed clips.
- Read the transcript before exporting. In Captions Cut every word is editable: tap it to fix the text or its timing.
How we checked
Vendors rank themselves first on benchmarks they choose, so their numbers are labelled as theirs and not compared across vendors. The independent figure is Artificial Analysis’s. Read on 22 October 2026.
Questions
Are automatic captions accurate enough to post without checking?
Rarely. Even at 2 to 5% WER a short clip usually has a mistake or two, often in a name. Read them first.
Which languages are transcribed best?
Widely spoken languages with lots of training data (English, Spanish, French, German, Japanese and others) score best. Captions Cut transcribes 53; less-resourced ones need more correcting.
Sources
- Artificial Analysis: ElevenLabs Scribe v2
- AssemblyAI: benchmarks
- Radford et al. (2022), Whisper
- ElevenLabs: speech to text
- Watanabe et al. (2020), CHiME-6
- Koenecke et al. (2020), PNAS
- YouTube: automatic captions
Each page read on .