Benchmark · November 13, 2026
English transcription, measured: 80 minutes of real meetings
We ran eight real English meetings — four people around a table, several of them non-native speakers, talking over each other — through JustTranscribe and through OpenAI's whisper-1, and scored both against the human transcripts, word by word. On meetings the two are level: 23.1% word error rate for JustTranscribe, 22.4% for whisper-1. JustTranscribe put the right speaker on 94.3% of the words and finished each ten-minute recording in about eight seconds.
On clean read speech whisper-1 is ahead (4.9% against 6.3%). We publish both results, the method and the data, because “99% accurate” means nothing without the recordings behind it.
Real meetings: the AMI Meeting Corpus
The AMI corpus is 100 hours of recorded meetings released by the University of Edinburgh and partners under CC BY 4.0, with human transcripts timed word by word. We took eight meetings from its three recording sites and used minutes 5:00 to 15:00 of each, from the single mixed recording of all the headsets, saved as a 64 kbps mono MP3 — the kind of file people upload.
| Meeting | Words | JustTranscribe WER | whisper-1 WER | Speakers found | Right speaker | Time (ours / whisper-1) |
|---|---|---|---|---|---|---|
| ES2002aEdinburgh, design-team scenario | 1,421 | 20.1% | 19% | 4 of 4 | 95.9% | 9 s / 35 s |
| ES2008aEdinburgh, design-team scenario | 1,493 | 21.4% | 22.4% | 4 of 4 | 95% | 7 s / 25 s |
| IS1001aIdiap (Switzerland), design-team scenario | 1,230 | 26.2% | 28.3% | 4 of 4 | 93.7% | 8 s / 28 s |
| IS1009aIdiap (Switzerland), design-team scenario | 1,439 | 21.7% | 24.9% | 4 of 4 | 92.7% | 6 s / 24 s |
| TS3003aTNO (Netherlands), design-team scenario | 875 | 18.5% | 16.8% | 3 of 3 | 97.4% | 8 s / 16 s |
| TS3005aTNO (Netherlands), design-team scenario | 1,089 | 29.7% | 22.9% | 4 of 4 | 94.3% | 7 s / 22 s |
| EN2001aEdinburgh, real project meeting | 2,085 | 15.6% | 15.4% | 4 of 5 | 96.1% | 6 s / 33 s |
| EN2002aEdinburgh, real project meeting | 2,256 | 30.9% | 28.1% | 4 of 4 | 90.9% | 10 s / 36 s |
| All eight (80 min) | 11,888 | 23.1% | 22.4% | — | 94.3% | 61 s / 220 s |
WER = substitutions + deletions + insertions, divided by the words in the human transcript. Lower is better. Bold marks the lower of the two.
Why a fifth of the words count as errors
Because of how people talk in meetings. Of JustTranscribe's 23.1 points, 16.2 are deletions — words in the human transcript that the tool left out — and only 6.3 are words it got wrong; whisper-1 shows the same pattern (16.9 and 4.6). The human transcribers wrote down every “yeah”, “okay” and half-started sentence, including the ones spoken underneath someone else. An automatic transcript of a single mixed recording follows the person who has the floor and drops most of what is said over them.
What that means for you: the transcript of a meeting reads as a clean account of who said what, and it is missing the cross-talk. If the overlapping remarks matter — a deposition, a negotiation — record each person on their own microphone, or expect to listen again at those moments.
Results vary by recording, which is why the table shows all eight and not an average alone: JustTranscribe had the lower error rate in three meetings and whisper-1 in five, and in one (TS3005a) whisper-1 was clearly ahead, 22.9% against 29.7%.
Speaker labels
JustTranscribe labels speakers as part of the same pass, with no setup. It found the right number of speakers in seven of the eight excerpts (in TS3003a that number is three: only three people speak in those ten minutes); in EN2001a five people speak and it found four. Among the words it transcribed correctly, 94.3% carried the right speaker (between 90.9% and 97.4% per meeting). whisper-1 does not label speakers, so it has no figure here.
Clean read speech: FLEURS
For the easy end of the range we used Google's FLEURS set (CC BY 4.0): 40 test sentences of US English, each read by a volunteer, joined into one 6.7-minute file after bringing every clip to the same loudness. whisper-1 scored 4.9% word error rate; JustTranscribe's engine scored 6.3% and 6.4% in two runs. This test counts “25” against “twenty-five” as an error for both, so both figures are higher than a reader would feel.
The same test in Latin American Spanish goes the other way — 2.0% for JustTranscribe against 3.0% for whisper-1 — which is where the product started. See Spanish audio to text.
Method
- JustTranscribe is the production pipeline as users get it on November 13, 2026: language detected automatically, no settings.
- whisper-1 is OpenAI's API model, told the language is English. Left to detect the language itself, it answered one meeting of accented English (IS1001a) in another language and scored 97%; we give it the hint so the comparison is fair.
- Scoring. Both outputs and the human transcript are lower-cased and stripped of punctuation; hesitation sounds (um, uh, mm, hmm) are removed from all three; numbers up to 99 are written as words; a short list of spelling pairs (okay/ok, colour/color, gonna/going to) is unified. Then a standard word alignment counts substitutions, deletions and insertions.
- Speaker accuracy. Each correctly transcribed word is compared with the speaker the human transcript gives that word; the tool's labels (Speaker 1, 2…) are matched to the real speakers by best overlap.
- Time is wall-clock for the transcription request, measured from Colombia.
- What this is not. Eight meetings and forty sentences, one day, two engines. No phone calls, no music, no other products — we only publish numbers for systems we ran ourselves on the same files.
Data and sources
- Download the results (CSV) — every row of the tables above, with substitutions, deletions and insertions. CC BY 4.0.
- AMI Meeting Corpus, CC BY 4.0 — groups.inf.ed.ac.uk/ami/corpus. Meetings ES2002a, ES2008a, IS1001a, IS1009a, TS3003a, TS3005a, EN2001a, EN2002a; Mix-Headset audio, manual annotations 1.6.2, minutes 5:00–15:00.
- FLEURS, CC BY 4.0 — huggingface.co/datasets/google/fleurs. en_us test split, first 40 sentences by file name.
- A 2.5-minute excerpt of ES2002a with JustTranscribe's unedited transcript is on the meeting transcript page.
To cite: JustTranscribe, “English transcription benchmark”, November 13, 2026, justtranscribe.ai/reports/english-transcription-benchmark. Questions or a correction: hello@justtranscribe.ai.