Transcription in academic research: the hours and dollars nobody budgets for

Open a finished qualitative dissertation and you rarely find three interviews, or three hundred. You find a couple of dozen. Mark Mason counted them: reviewing 560 doctoral dissertations that used qualitative interviews, he found a mean of 31 interviews and a median of 28, with samples ranging from a single interview to 95.

The distribution has a shape worth knowing. The most common sample sizes Mason recorded were 20 and 30 interviews, followed by 40, 10 and 25 – so the typical thesis is processing somewhere between 20 and 40 recordings. That is the real unit of work this piece is about: not one file, but a cohort of them, arriving over a season of fieldwork. If you are staring at that cohort now, our guides to qualitative research transcription and transcribing an interview for a dissertation are the practical companions to this one.

Why 28 and not 8? Because qualitative researchers keep interviewing until new interviews stop teaching them anything new – the point called saturation – and saturation tends to arrive later than beginners expect. In the study that made the question famous, Greg Guest and colleagues analysed 60 in-depth interviews and found the main themes were saturated within the first twelve, with the basic elements of those themes visible as early as six.

Twelve sounds reassuringly small until you read the follow-up work. Monique Hennink and colleagues drew a sharper line: analysing 25 interviews, they found code saturation – hearing the full range of issues – at nine interviews, but meaning saturation, a full understanding of each issue, took sixteen to twenty-four. A study that wants to understand each theme, not just list it, lands squarely back in the dozens. Which is exactly why the median thesis looks the way Mason found it.

Now convert interviews into hours, because hours are what you actually transcribe. A single qualitative interview commonly runs anywhere from 30 minutes to several hours. Take the conservative middle – call it an hour each – and a 28-interview study is roughly 28 hours of audio. A study at the upper end of Mason's common range, 40 interviews at 60 to 90 minutes apiece, is 40 to 60 hours.

The honest planning number, then, is something like 20 to 40+ hours of recorded audio for a normal qualitative thesis. That is the pile sitting on your drive after fieldwork. Everything downstream – transcription, coding, pulling quotes – is priced against that pile. And a pile that size is exactly where consumer tools start to strain, which is the subject of our guide to transcribing long audio.

Transcribing by hand is slower than people remember. A University of Bath research-data guide puts it at 4 to 7 hours of work to transcribe a single hour of audio, and names the main disadvantage of doing it yourself in one word: time-consuming.

Run that against the pile. Twenty-eight audio-hours at the optimistic end of Bath's range – four hours of typing per hour of tape – works out to about 112 hours of transcription. That is roughly three full working weeks spent at the keyboard before a single line of analysis begins. We took the timing question apart on its own in how long it takes to transcribe an hour of audio; the short version is that the ratio rarely dips below four to one, and rises sharply once you need strict verbatim conventions.

The obvious escape is to pay someone else, and here the price is legible because it is charged by the minute. Rev, one of the larger transcription vendors, lists human transcription at $1.99 per audio minute – about $119 per audio hour – and AI transcription at $0.25 per minute. (Last verified: 2026-07-13; vendor prices move, so re-check before you budget.)

Do the multiplication on a median study. Twenty-eight audio-hours of human transcription at $1.99 a minute works out to roughly $3,340 – for one thesis, before revisions or a second pass. The AI rate is far gentler on the wallet. But the thing that catches researchers out is not the per-minute number at all; it is the way that number is usually packaged, which the next two sections take up. We keep a running teardown of the market in how much transcription costs and the 2026 transcription cost data.

There is a second bill that never appears on an invoice: what your ethics board requires of the audio itself. Interview recordings are treated as sensitive by default. Washington State University's human-research program states plainly that audio and video recordings are always considered Confidential data, and that a personal smartphone should not be used to record participants at all.

Where the file is allowed to live is governed too. The University of Connecticut's IRB requires that University-managed devices be used to collect voice recordings, bars personal laptops and phones for identifiable or confidential data, and notes that federal rules require records be kept for at least three years after the research is complete – a floor, not a ceiling, with some institutions specifying longer.

None of that is compatible with quietly uploading interviews to whatever free tool ranks first in a search. It is the reason the question "where does my audio go, and does anyone train a model on it" is not paranoia but paperwork. We wrote the practical version up as confidential transcription.

Set the ethics aside and there is still a structural mismatch between how transcription is sold and how a thesis is written. A thesis is a burst: 20 to 40 hours of audio arrive over a few weeks of fieldwork, and then the tap shuts off for months of coding and writing. Most paid transcription is sold as a subscription – a fixed block of minutes that refills every month and expires if you do not use it.

That shape fits a newsroom or an agency with steady weekly volume. It fits a dissertation badly. You either overpay for months of minutes you never touch, or you do the dance every bursty user eventually learns: subscribe for one month, push everything through, cancel before it renews. We heard that exact routine again and again in 100 Reddit threads about transcription. What the work actually calls for is pay-per-use – buy the minutes the thesis needs, keep the ones it does not. That instinct is why transcription without a subscription is a search people make on purpose.

Stack the four constraints and the right tool for a thesis nearly specifies itself. It has to swallow a 20-to-40-hour pile without capping or dying on the long files. It should charge for what you use, not rent you minutes by the month. It has to give you the actual transcript – with speakers separated, since an interview has at least two – in a format your analysis software will read. And it has to be able to answer the ethics-board question in writing.

That is the shape we built Pepys to fit: pay-once credits that never expire, no per-file length cap, real exports, speaker diarization so a two-person interview comes back as two people, and no training on your audio, ever. None of that is a pitch so much as a description of what the fieldwork already needed. If you want to see the per-minute cost against your own pile, the pricing is one page, and you can start a transcript with a single interview to test it before you trust it with the whole cohort.