How Long to Transcribe 1 Hour of Audio?
About four hours by hand and anywhere from minutes to roughly an hour by machine. A sourced table, our own measured speeds and what changes them.
The Scriptorium team ·
By hand, one hour of audio takes about four hours for an average typist and two to three for a professional, according to Rev's page on transcription time. A machine is faster but not instant: on our own CPU-only server, the Whisper small model processed English at a real-time factor of 1.25, which is about 75 minutes for an hour of audio, while tiny needed about 15. A graphics card or a quiet, faster computer would be quicker; we have not measured those.
The numbers, side by side
| Method | Time for 1 hour of audio | Where the number comes from |
|---|---|---|
| A person typing (average) | about 4 hours | Rev's page, checked 2026-10-26 |
| A professional transcriptionist | 2 to 3 hours | Rev's page, checked 2026-10-26 |
| Whisper tiny, English, our server | about 15 minutes | Real-time factor 0.24, our measurement |
| Whisper base, English, our server | about 32 minutes | Real-time factor 0.53, our measurement |
| Whisper small, English, our server | about 75 minutes | Real-time factor 1.25, our measurement |
| Whisper small, Arabic, our server | about 76 minutes | Real-time factor 1.27, our measurement |
The real-time factor is processing time divided by recording length: 0.5 means a 10-minute recording took 5 minutes. The hour figures above are those factors multiplied by 60, so they assume speed holds steady over a longer file, which we did not test on a real hour.
How we measured, and why not to over-read it
- Hardware: 4 virtual CPUs with no graphics card, capped at 150% of one CPU-second per second (cgroup cap). The same server was busy with other work (load average between 9.4 and 24.4), so a quiet machine would have been quicker.
- Clips: 6 English clips (487 seconds of audio) and 3 Arabic clips (246 seconds), measured on 2026-10-24. One run per model, no confidence intervals.
- Software: faster-whisper with int8 on the CPU. The project says its implementation is "up to 4 times faster than openai/whisper for the same accuracy" (faster-whisper README, checked 2026-10-26).
- A steadier measure than wall-clock time is CPU time: English small used 0.78 CPU-seconds per second of audio, Arabic 1.36, on our clips.
All the numbers, the clips and the method are on the measurements page.
What makes a transcript slower or faster?
- Audio quality. Rev says the recording's quality has a major impact on transcription speed, and that noise, accents and switching between two languages make it harder for a human transcriber too.
- Model size. A bigger model is slower. On our clips the smaller models were quicker and made more errors; see the table below.
- Language. In our measurement Arabic used more CPU time per second of audio than English with the same model. Three clips are not a rule, but it is a reason to allow longer for Arabic.
- Hardware. A GPU, or simply a quiet computer, changes the speed far more than anything you do to the file.
- A queue. On a shared service you also wait for the jobs ahead of yours (see below).
Faster is not free: speed against errors
| Model | English: hour takes | English word error rate | Arabic: hour takes | Arabic word error rate |
|---|---|---|---|---|
| tiny | about 15 min | 7.5% | about 26 min | 64.4% |
| base | about 32 min | 5.2% | about 38 min | 60.2% |
| small | about 75 min | 3.7% | about 76 min | 31.8% |
Word error rate is the share of reference words the draft got wrong (what that means). These figures come from short, clearly read clips (English audiobook speech; Arabic read by Egyptian speakers), and the Arabic set is only three clips, so they say how the models compare, not what your recording will score. On the Arabic clips the smallest model's error rate was roughly double the small model's, which is why we do not recommend tiny for Arabic.
And the time you still spend afterwards
A draft needs a check, and checking against the audio means listening, so budget at least the length of the recording again if the text matters. For a quote, a name or a number, listen to that part rather than trusting that the draft reads well; mistakes in a transcript tend to be fluent, which is what makes them easy to miss.
What this means on our free service
The no-account preview takes one file at a time, up to 15 minutes and 25 MB, with 3 tries a day, so it is a way to test a short clip, not to transcribe an hour. A free account includes None transcription minutes a month. All jobs go through one worker shared by everyone, so the wait depends on how many jobs are ahead of yours; we do not promise a turnaround time. If you have an hour-long file, see how to transcribe a one-hour recording for free and how to transcribe a long recording.
How to shorten the wait
- Send audio, not video. A video file is mostly picture; extract the sound first and upload that.
- Cut the silence and the music at the start and end.
- Split a very long file at natural breaks so you can check and use the first part while the rest runs.
- Pick the smallest model that is accurate enough for your language, and test a one-minute clip first.
Checked on 2026-10-26. Rev's page on how long human transcription takes (average person about 4 hours per audio hour, professionals 2 to 3 hours, quality factors); the faster-whisper README (the 4x speed claim, which is the project's own). Rev sells transcription, so its figures are a vendor's estimates, not a measured study. We found no neutral measurement and cite none. Not read: any vendor's automatic-transcription turnaround time. Menus, prices and limits change; confirm on the vendor's page before you rely on them.