How long does it take to transcribe an hour of audio?

Plan for about four hours of work to transcribe one hour of clear audio by hand. That is the 4:1 ratio the transcription industry quotes almost universally, and it is the single most useful number to anchor a project on. A three-hour set of interviews is closer to a day and a half of typing than an afternoon.

That figure is the average, not the floor and not the ceiling. A fast professional on clean audio can beat it; verbatim work on a noisy, multi-speaker recording can take more than double it. And it describes a person at a keyboard - not the calendar time a transcription service quotes, which is measured in days, nor an AI model, which is measured in seconds. This piece walks all three, so you can plan the right one.

The ratio is not folklore. It falls out of a smaller, concrete measurement: most people need about an hour to transcribe just fifteen minutes of clear, slow audio, once you account for stopping, rewinding, and fixing what you misheard (Rev, last verified 2026-07-12). Multiply that out and an hour of audio is four hours of work. Four fifteen-minute blocks, four hours.

Note the word "clear." That estimate assumes a single speaker talking at a reasonable pace with good sound. The moment you add crosstalk, an accent you have to lean into, or a phrase you replay five times to get right, the fifteen-minutes-per-hour pace slips, and the whole ratio climbs with it.

The 4:1 average hides a wide range, and knowing the range is what stops you underbooking a project. On the fast end, a skilled professional transcribes clean audio in roughly two to three hours per audio hour, and the very best can put down thirty minutes of audio in an hour of work - a 2:1 pace (Rev, last verified 2026-07-12). One vendor pegs standard work at three to four hours per audio hour, or about fifteen to twenty minutes of audio cleared per working hour (Ditto Transcripts, last verified 2026-07-12).

On the slow end, difficulty stretches everything. Complex audio pushes a professional toward six hours per audio hour, and true verbatim - every false start, every filler word, sometimes visual detail too - runs higher still. Vendors frame four hours as a minimum that can "easily reach" ten as difficulty rises (Verbit, last verified 2026-07-12).

This is where it stops being vendor marketing. Peer-reviewed qualitative-research guidance lands in the same place: transcription takes at least three hours per hour of talk, and up to ten hours per hour for a fine level of detail (Bailey, 2008). A 2024 methods paper in the European Journal of Cardiovascular Nursing gives the identical estimate - three to ten hours to transcribe one hour of raw data - and calls it a laborious, time-consuming task. When independent researchers and commercial services converge on the same spread, treat it as real.

If four hours of typing sounds manageable, watch the gap between doing the work and receiving the work. When you hand a file to a professional service, the clock includes queueing, a review pass, and proofing on top of the raw transcription - so even a four-hour job can take around forty-eight hours to come back (Verbit, last verified 2026-07-12).

Across file lengths, human-delivered turnaround typically runs from one business day (with accuracy not guaranteed at the fastest tier) to six business days for files three hours or longer (Verbit, last verified 2026-07-12). None of that is wasted; it is the cost of a checked, human-quality transcript. Just budget the calendar, not only the labor - and if the calendar is what you are weighing against the bill, our guide to how much transcription costs puts the two side by side.

Automatic speech recognition changes the unit. Instead of hours per audio hour, you measure seconds. Published benchmarks make this concrete: running Whisper Large v3 on Groq hardware clocks a 164x real-time speed factor in independent testing, transcribing a ten-minute file in 3.7 seconds - which implies a one-hour recording finishing in roughly twenty-two seconds (Groq / Artificial Analysis benchmark).

You do not need a datacenter to beat real time, either. In SYSTRAN's published faster-whisper benchmark, thirteen minutes of audio transcribes in 59 seconds on a GPU (int8, large-v2), against 2m23s for the reference implementation. Even on a plain laptop CPU, the same thirteen minutes takes about 1m42s with a smaller model - still faster than the audio itself plays.

These are published benchmarks from vendor and open-source sources, not a Pepys first-party test. But the shape is consistent everywhere: an hour of audio that a person spends a working day on, a model finishes before you have refilled your coffee. Speed is not the same as accuracy - that is a separate question we cover in how accurate AI transcription is - but on raw time, it is not close.

The near-instant numbers come with a condition worth stating plainly: AI speed is hardware- and model-dependent. Groq's 164x is a cloud-GPU result, and the faster-whisper figures above use a datacenter GPU or a deliberately smaller model. Change the setup and the speed changes.

Run the largest Whisper models on an ordinary CPU with no GPU and the picture can flip: the sub-minute CPU figures above are for a deliberately smaller model, and without GPU acceleration the heaviest models slow down sharply and can drop below real time. So "AI is instant" is true on capable hardware and with a right-sized model, and misleading if you point the heaviest model at a modest laptop. The practical takeaway: on a cloud service or a decent GPU, an hour of audio is a sub-minute job; on your own machine, pick a smaller model or expect to wait.

Turn the ratios into a day's work. Assume a focused eight-hour working day. At the 4:1 average, that day clears about two hours of audio. A fast professional at a 2:1 pace clears roughly four. On verbatim or difficult material closer to 8:1 or 10:1, you will not finish a single audio-hour in a day. That is the honest planning envelope for transcribing by hand.

Hand it to a service and the constraint moves from labor to calendar: one to six business days per batch, depending on length and tier (Verbit, last verified 2026-07-12). Send it to AI and the transcription step effectively disappears - your real limit becomes upload speed and file handling, not the model's clock. A hundred hours of interviews that would be weeks of typing becomes an afternoon of uploading.

Two things make the AI route feel slower than the benchmark: getting long or oversized files in at all, and untangling who said what. Both have known fixes - see how to transcribe long audio for the chunking approach that keeps a two-hour file from dying halfway, and how to transcribe multiple speakers if your recording needs labeled turns.

Transcribe by hand when the slowness is the point - forensic verbatim, or a researcher who wants the deep familiarity that comes from typing every word. Use a human service when accuracy has to be guaranteed and a few business days is fine. Reach for AI when speed is the constraint and a review pass on the output is acceptable, which is most first drafts, most interviews, most episodes.

If privacy or budget rules, the honest answer is often free and local: open-source Whisper (via tools like faster-whisper) runs on your own machine, never uploads, and costs nothing but compute - and as the benchmarks show, even a laptop can beat real time with a right-sized model. There is no shame in that being the best fit.

When you want the fast route without running the model yourself, that is where Pepys sits: upload an hour of audio, get the transcript back in the time it takes to read this paragraph, and pay once for the minutes you use rather than renting a subscription. You can point it at a file directly with the audio-to-text tool, and the pricing is usage-based by design. It is one honest option for the fast lane, not the only one - and if you want the texture behind why people go looking for a better tool in the first place, we read 100 Reddit threads about transcription and wrote down what everyone hates.