Most transcription is finished when it reads well. Academic transcription is finished when someone else could check your findings against it. That difference decides almost everything: what detail you keep, how you label people, who's allowed to hear the recording, and what you can say in your methods section afterwards.

By academic transcription I mean turning research recordings, such as interviews, focus groups and seminars, into text under rules your study can defend. If you're about to hire a transcriber or try an AI tool, here's what I'd pin down first. We run a human transcription service and a self-serve tool, so I'll also say where each one fits and where it doesn't.

Why academic transcription is different

In research, a transcript is data. The quotes in your thesis point back to it, your coding rests on it, and an examiner or co-author may ask how it was made.

That's why research data archives care about the paperwork. The UK Data Service says each text file should contain enough context to stand on its own, with at least a unique identifier, the date of collection and the setting where relevant. It also says "Transcription conventions must be documented." Settle your conventions before you start, so the transcripts stay comparable from the first interview to the last.

What to put in the brief

A good brief is short. It answers four questions a transcriber would otherwise answer for you.

How much detail do you need?

The usual choice is between full verbatim, with every false start and filler, and clean verbatim, which drops them without changing the meaning. We have worked examples of clean verbatim if you want to see the difference on the page. The right answer follows from your method. Conversation analysis and discourse work usually need the detail. Thematic work usually doesn't.

How should speakers and time be marked?

Use fixed labels such as Interviewer and P1, and say whether you want a time code on every turn or at set intervals. If you'll quote from the transcripts, ask for numbered lines too, because they make a quote easy to cite.

What happens when nobody can hear a word?

Tell the transcriber to mark it and move on. [inaudible 00:12:40] is honest. A confident guess is how errors get into your data.

Who removes names, and when?

Some projects replace names during transcription, others afterwards. Either works if it's decided up front. If you're doing it yourself, our free interview anonymizer swaps names for pseudonyms in your browser, and nothing is uploaded. Keep the key that links pseudonyms to people somewhere else.

How accurate does academic transcription need to be?

Accurate enough that no finding depends on an error. That's a higher bar than any percentage, and the research shows why.

A 2023 study in JMIR Mental Health measured transcription services on 65 research interviews against a reference the research team produced themselves. The best automatic service had a median word error rate of 8.9%, and a commercial human service had 7.6%. The authors concluded the gap may be narrowing. I'd draw a second conclusion: the human transcripts weren't flawless either, so no source lets you skip the checking.

Then there's the question of which errors. Researchers at CISPA compared eleven providers on recordings about cybersecurity and found AI services often distorted the meaning. "Hashes" became "ashes". In that field, that's a different sentence. Their test dates from December 2022, so the tools have improved, but an error rate still can't tell you which word went wrong.

So use error rates to shortlist a tool, not to skip the work. Listen to the recording for every passage you quote, and correct names, negations and numbers first, because those are the slips that change what someone said.

Who is allowed to hear the recording?

The rules differ by institution and country, so take what follows as questions for your ethics board or data protection officer, not as advice on what's permitted.

Three examples show the range. The University of Jyväskylä says that a researcher sending personal data to an external transcription company must draw up a data processing agreement, signed by the head of department. The University of Canterbury in New Zealand doesn't recommend AI meeting transcription tools for sensitive or in-confidence research data, and points researchers to Microsoft Teams and an offline service first. Loughborough University warns that some services act as data controllers, and that many use uploaded audio to train their models.

Ask before you record, because your consent form has to match what actually happens to the audio. For any provider, I'd want answers to these:

  • Where is the audio processed and stored?
  • Is it used to train models?
  • How and when is it deleted?
  • Can they sign a data processing agreement?
  • Which people can listen to it?

Institution tool, AI tool or human transcriber?

Start with whatever your institution already provides, since that's where approval is most likely to be in place. If there's nothing, or it can't handle your language, you have two real options.

A self-serve AI tool suits lower-risk recordings and projects where you'll correct the text yourself anyway. Ours is Spirelight Transcription. It handles Danish and many other languages, and you can code interviews in the same project. Know the limits before you use it. Transcription runs on EU servers, files and accounts are stored in the US, and standard accounts don't include a data processing agreement, though organisations can ask us about an Enterprise agreement. For identifiable participants that may rule out a standard account, and your institution should make that call.

A human transcriber suits sensitive material, difficult audio, several speakers at once, or languages where the tools fall short. It costs far more. In the JMIR study the human service cost US $1.50 a minute and the best automatic one US $0.024. For a handful of interviews that carry a whole thesis, that can still be money well spent.

If your project needs human transcripts in another language, or to a written spec your supervisor can inspect, tell us what you have. Our transcription service covers 13 languages through native transcribers, and you can start with a free sample.