Verbatimly

We tested Verbatimly on a real TED talk.
Word error rate: 1.9%

Most transcription tools say "highly accurate" and leave it there. We wanted a number you can check, so we ran a public talk with a human-written transcript through Verbatimly and compared every word.

1.9%Word error rate
1,502Words in ground truth
29Errors total
4sTo transcribe 9:15

What we tested

Judson Brewer's TED talk A Simple Way to Break a Bad Habit — 9 minutes 15 seconds, one speaker, natural pace, spoken to a live audience rather than read from a script.

TED publishes an English transcript written and reviewed by humans. That transcript is the answer key. We uploaded the audio to verbatimly.ai, took the output, and aligned the two word by word.

Why a TED talk and not a clean studio recording: because you are not transcribing studio recordings. Real speech has filler words, false starts and uneven pacing. A benchmark that hides those is not measuring the thing you care about.

The numbers

MetricValue
Words in TED's human transcript1,502
Words in Verbatimly's output1,513
Substitutions12
Deletions3
Insertions14
Word error rate1.9%

The 29 errors were concentrated in proper nouns and contractions. Sentence structure held throughout — the transcript reads as prose, not as a word soup.

How to check this yourself

  1. Download the audio of any TED talk that has an official transcript. TED offers both on every talk page.
  2. Upload the audio to verbatimly.ai. No account needed.
  3. Normalize both texts the same way: lowercase, strip punctuation, keep apostrophes as part of the word.
  4. Align them with word-level edit distance and count substitutions, deletions and insertions.
  5. Divide by the word count of the answer key. That is your WER.

If your number comes out far from ours, send us the file. A result that only holds on our machine is not a result.

Cite this

Verbatimly accuracy test (November 2026). 9-minute TED talk, official human transcript as ground truth. 1,502 words, 29 errors, WER 1.9% (98.1% word accuracy). https://verbatimly.ai/accuracy
Try it on your own recording →