Transcription speed and accuracy, measured

We ran the Whisper models that power TryScriptorium (tiny, base and small) on public test clips, on our own server. On 487 seconds of English speech the small model's word error rate was 3.4% on clean speech and 4.0% on harder speech, and 31.8% on 246 seconds of Arabic. These are measurements on specific clips, not a promise about your recording.

Last updated: 24 October 2026

Measured on 24 October 2026. Every number is real and stored in a downloadable data file: transcription-speed.json.

Results: English

WER is word error rate (lower is better). "Clean" is LibriSpeech dev-clean; "harder" is dev-other, which has more difficult speakers and recordings. Real-time factor is wall-clock processing time divided by audio length.

Results: Arabic

Read Arabic sentences from FLEURS (ar_eg, Egyptian speakers reading standard written text), scored after Arabic letter normalisation (diacritics removed; alef, yeh and teh-marbuta forms unified).

What the speed means

The server was busy with other work during the run (1-minute load average between 9.4 and 24.4 on 4 vCPUs) and the benchmark was limited to 150% of one CPU-second per second (cgroup cap), so wall-clock times are slower than an idle machine would give. The better figure is CPU-seconds per audio-second. The small model used 0.78 CPU-seconds per second of audio, about 47 CPU-seconds for each minute of recording.

Method

• Hardware: Intel Core Processor (Broadwell, no TSX, IBRS), 4 vCPUs, 7.7 GB RAM, no GPU. This is the same server that runs the website, so it was not idle.
• Software: faster-whisper 1.2.1, CTranslate2 4.8.2, Python 3.14.4.
• Settings mirror the production engine: CPU, int8, beam size 5, voice-activity filter on. The language was set explicitly (no auto-detection).
• Data:
  • LibriSpeech dev-clean and dev-other: English read audiobook speech (CC BY 4.0), built from public-domain LibriVox recordings; 3 clips from each set.
  • FLEURS ar_eg (test split): Read Arabic sentences by Egyptian speakers (CC BY 4.0); 3 clips of consecutive sentences.
• Clips: 6 English and 3 Arabic clips of roughly 75-100 seconds, built by joining consecutive sentences, each with its reference text.
• WER: word-level edit distance divided by reference word count. English: lower-cased, punctuation removed, no number normalisation (so "10" against "ten" counts as an error). Arabic: diacritics removed, letter forms unified.
• CPU time is the process's user plus system CPU time while transcribing. Memory is peak resident size of a separate process per model.
• Repeats: one run per model, no confidence intervals.

What these numbers do not show

• Only clearly recorded read speech. Phone calls, noise, crosstalk and strong accents give higher error rates.
• Only English and Arabic. We did not measure other languages.
• The medium and large models were not run: the shared production server cannot safely fit them while serving the site.
• Few clips, one run: small differences between two models are not evidence.
• Arabic scoring is strict: spelling variants (for example hamza forms) and dialect words count as errors even when a reader would accept them, so the Arabic error rates are not directly comparable with the English ones.
• This is not a promise of accuracy for your audio. The free preview uses the small model.

How to cite

TryScriptorium. "Transcription speed and accuracy, measured." 24 October 2026. https://tryscriptorium.com/research/transcription-speed

See also the fact sheet and Arabic transcription.