Accuracy and speed, measured
The production engine on public reference audio, as served, with the normalisation and the method stated.
What was run
WhipScribe production engine, both tiers as served, on 70 runs, 2026-09-25. Audio: FLEURS test split (CC BY 4.0), read speech, clips joined into ~5-minute files; LibriSpeech test-clean / test-other (CC BY 4.0), read audiobook speech. Files were submitted through the same upload path customers use, four at a time, while customer jobs ran; nothing in the engine was changed for the run. One engine serves every job, and that is the one measured. Normalisation: NFKC, lowercase, punctuation removed, whitespace collapsed; CER for Japanese, Chinese and Korean (no reliable word boundaries), as in the Whisper paper; wer_ex_numbers drops tokens containing digits from both sides.
Read speech is easier than conversation. These numbers are a floor for clean audio, not a promise for a phone call in a café; the hallucination study covers the failure modes on real recordings.
Results
| Language | Set | Files | Minutes | Error rate, median | Worst file | Speed |
|---|---|---|---|---|---|---|
| Cebuano | read | 5 | 26.4 | WER 47.5 % | 57.5 % | 21.1× |
| Chinese (Mandarin) | read | 5 | 25.9 | CER 9.8 % | 12.7 % | 26.7× |
| German | read | 5 | 25.8 | WER 8.8 % | 11.8 % | 28.7× |
| English | read | 5 | 26.3 | WER 11.2 % | 14.7 % | 30.6× |
| English | test-clean | 5 | 26.5 | WER 7.0 % | 10.3 % | 29.2× |
| English | test-other | 5 | 26.8 | WER 10.7 % | 22.3 % | 27.3× |
| Spanish | read | 5 | 26.4 | WER 4.5 % | 13.2 % | 28.9× |
| Filipino (Tagalog) | read | 5 | 26.0 | WER 21.0 % | 36.1 % | 26.3× |
| French | read | 5 | 26.6 | WER 8.0 % | 14.1 % | 23.5× |
| Hindi | read | 5 | 26.4 | WER 20.7 % | 25.0 % | 17.2× |
| Japanese | read | 5 | 26.5 | CER 9.8 % | 16.3 % | 29.8× |
| Korean | read | 5 | 26.3 | CER 11.4 % | 18.1 % | 29.1× |
| Portuguese | read | 5 | 26.3 | WER 4.5 % | 8.7 % | 30.6× |
| Russian | read | 5 | 26.2 | WER 6.5 % | 11.2 % | 28.9× |
Error rate is word error rate, or character error rate for Japanese, Chinese and Korean. Speed is audio duration divided by transcription processing time, median per group, on a shared production GPU. End to end, from the upload request to the finished transcript being available, the median five-minute file took 46 seconds.
Reproduce it
The test sets are public (FLEURS and LibriSpeech, CC BY 4.0). Join the test-split clips into five-minute files in manifest order, upload them, score with the normalisation above. The corpus assembly and scoring scripts are available on request at contact@whipscribe.com.
Related measurements
How often the engine invents text on real recordings, by language: hallucination study. What people send in a month: what people transcribe. Method: how we test.
See also: Transcription services · Turnaround time · Evidence log · Pricing · Privacy · Terms