Report, measured October 20, 2026

Speed and accuracy, measured

Two questions people ask before they upload: how long will it take, and how many words will be wrong? This page answers both with numbers we measured, not numbers we picked. Speed comes from 3,836 real customer files. Accuracy comes from 1,520 transcriptions of public test recordings that anyone can download, and every transcript we scored is published below.

Speed on real files

Every transcription that finished between September 20 and October 20, 2026, excluding our own test accounts. "Wait" is the time from submitting the file or link to the finished transcript, which is what you experience. "Processing" starts when the job leaves the queue. For links it includes fetching the recording.

• "n/a" means fewer than 20 files in that group, too few for a fair 9 in 10 figure.
• Left out of the tables: files re-run after they first finished (34), files finished by a backup engine (31), and files whose processing started more than 6 hours after they were submitted (51). For those, our records cannot show how long anyone waited, because a later re-run of a file, for example after an upgrade, restarts its clock.

How long a file can be

• The longest file completed in the window above was 17 h 15 min of audio, transcribed end to end.
• Paid plans and passes: uploads up to 5 GB per file, with no separate length cap.
• Free account: uploads up to 1 GB and 30 minutes per file. Your first file can run up to an hour.
• Without an account: uploads up to 100 MB, first 30 minutes transcribed.
• Links: recordings up to 24 hours long on paid plans and passes, up to 3 hours without one.
• Our premium engine, with speaker labels, takes files up to 10 hours. Longer files are still transcribed, on our backup engines.

Accuracy on public test sets

Word error rate (WER) is the share of words a transcript gets wrong: words swapped, missed or added, divided by the words actually spoken. Lower is better. 5% means about 1 word in 20 needs a fix.

Clips that returned an error: premium engine 0, standard engine 0. Each set is scored as one pool of words, so long clips count for more than short ones.

Method

• Test sets. LibriSpeech test-clean and test-other (public domain audiobooks, CC BY 4.0 corpus), 100 clips each, and FLEURS test split (CC BY 4.0), 80 clips per language. Clips were picked at random with a fixed seed (20261020), before any transcription ran.
• Engines. Each clip went through the same engine, settings and language choice that production uses for that tier: the premium engine that runs paid files, and the standard engine that runs free files past the monthly premium allowance. The language was set to the clip's language, as when you pick it in the uploader.
• Scoring. WER computed with the open source jiwer library. Before scoring, reference and transcript get the same cleanup: for English, the standard English text normalizer used in public speech research (lowercase, punctuation removed, numbers and spellings unified); for other languages, lowercase and punctuation and symbols removed.
• Check it yourself. Every clip ID, reference text and transcript we scored is in per-clip-results.json, and score.py recomputes every number in the table from it.

Limits, stated plainly

• These test sets are read speech: one speaker, a decent microphone, little noise. Real meetings, phone calls, crosstalk, heavy accents and music raise the error rate for every engine. Treat these numbers as the best case, and test your own audio.
• 760 clips per engine is a fair sample, not a census. Differences of a point or so between languages are within the noise.
• Hindi and Arabic score worse partly because of spelling: Hindi words with more than one accepted spelling, and Arabic vowel marks that the reference writes and the transcript leaves out, count as errors here even when a reader would accept them. You can see each case in the per-clip file.
• We tested 7 languages. We support many more, and we have not measured them here.
• Speed depends on load. These are the numbers for the window above, and we will re-run them.
• We run this benchmark on ourselves, which is why we publish every transcript it scored.