How to Transcribe an Interview

When it comes to qualitative research, your transcripts are your data. Every theme you report, every quote in your findings chapter and every claim your examiners test traces back to them, so a transcript nobody checked puts the whole analysis on a shaky foundation. And the tools that promise to do the job in minutes are very good at producing text that looks right at first glance, but isn’t.

Having coached 10,000+ students, we’ve found the problems rarely start at the coding stage. They start here, with a transcript in the wrong style, an automated draft nobody listened back to, or an original that got edited when it should have been kept. In this post, I’ll walk you through how to transcribe an interview properly: which style your analysis method requires, the six key steps involved, and examples of real transcripts.

---

Before anything, check your ethics approval

Before you open a transcription tool (and ideally, before you even hit record), reread your ethics approval and the consent form your participants signed. In the US, that’s your IRB protocol; elsewhere, it’s your ethics committee’s approval. I’ll call it your protocol from here on.

Transcription is data handling, and it’s covered by the same promises you made about who could hear the recordings and where they’d be stored, and in many countries by data protection law as well. The GDPR in the EU, and its UK equivalent, treat any information relating to an identifiable person as personal data, and a recorded interview is exactly that.

The rules vary by institution, so the homework step is yours. Some universities vet transcription services in advance: CSU Bakersfield’s IRB, for instance, lists services it has already approved and asks researchers using any other service to include that service’s data-security page in their protocol. Others say only that whatever platform you use has to be described in your submission.

Three questions you need clear answers to before you start:

1. Where may the audio go? A cloud tool uploads your recordings to someone else’s servers. If your consent form promised the recordings would stay with the research team, that upload may break the promise.
2. Who may hear it? A human transcriber, yours or a service’s, is another person hearing the interview. Many protocols allow it under a confidentiality agreement; some don’t allow it at all.
3. When must it be deleted? Transcription creates copies: with the transcription service provider, on your hard drive, with your email provider, and so on. Your data plan’s deletion date applies to all of them.

If your protocol doesn’t mention transcription, ask your IRB or ethics committee before you upload anything, not after. Amending an approval is paperwork. Explaining an unapproved upload is a much harder conversation.

---

Verbatim or intelligent verbatim?

There are two main styles of research transcript, and the right one is decided by your qualitative analysis method, not by what’s easiest to type.

• Verbatim (the literature calls it naturalized transcription) records speech in as much detail as possible: every “um”, repeated word, false start and pause. Oliver, Serovich and Mason note it’s most often seen in conversation analysis, where how people speak is the data.
• Intelligent verbatim (denaturalized transcription) keeps every word of substance but drops the fillers, repetitions and false starts. The aim is accuracy about meaning rather than about sound. The same paper connects it with grounded theory, ethnography and critical discourse analysis.

Here’s the difference on real speech. The excerpt comes from our own podcast, where a dissertation chair answers a question about milestones, first as spoken and then cleaned up:

Notice what the second version keeps. “There wasn’t as many” stays, because correcting a participant’s grammar is changing their words, and “kind of” stays because it softens the claim. Intelligent verbatim removes noise, never meaning. If you’re unsure whether something is noise, keep it.

For most thematic analysis, intelligent verbatim is enough. If your method treats hesitation, emphasis or turn-taking as meaningful, as conversation analysis does, you need full verbatim, and your advisor (or supervisor, in the UK) may expect a notation system for pauses and overlaps. Whichever you choose, state it in your methodology chapter, because it changes what your quotes can be used to argue.

---

How to transcribe your interviews

Transcribing an interview or focus group recording involves six steps. The order matters: the checking in step 4 only works while the interview is still fresh in your memory, and the de-identification in step 5 has to happen on a copy.

1. Get audio worth transcribing

Most transcription problems are recording problems. Record in a quiet room, test the setup before the first interview, and run a second device as a backup. Make sure you end up with a digital audio file you can play back, pause and share with whoever helps you transcribe. An analog tape recorder won’t give you one. There’s more on the interview itself in our post on common interview mistakes.

Our channel has a short walkthrough of the same mistakes:

2. Produce a first draft

You have three routes, and each has a place.

By hand. You listen and type, pausing constantly. It’s slow, but it’s the deepest early reading of your data you’ll ever do. It also keeps the audio solely on your computer, simplifying any data-sharing compliance requirements.

Automated, then corrected. Your meeting platform (Zoom, Microsoft Teams or Google Meet), a dedicated app such as Otter.ai, or a speech model produces a draft, which you then correct line by line. Some open-source models, such as OpenAI’s Whisper, run entirely on your own machine, which helps if your protocol rules out cloud uploads. If you plan to use your meeting platform’s built-in transcription, check the setting before the interview: in several platforms, recording and transcription are separate switches.

Handed off. A transcription service does the typing. This is normal and typically allowed, as long as your protocol permits another person to hear the audio. The analysis, and every interpretation in it, stays yours. If you go this route, ask the service provider what exactly they do with your files (where they’re stored, who can access them and when they’re deleted), and get the confidentiality terms in writing.

3. Format it for analysis

A transcript is a working document, so format it for the work you’ll do on it. This is especially important as your data set grows in size: a consistent, rich structure saves you a lot of pain in the long run. Start each file with a header: the participant’s pseudonym, the interviewer, the date, the setting and the recording’s length. Then:

• Label every speaker consistently. “I” for interviewer and “P1”, “P2” for participants is common and anonymous by design.
• Add timestamps at every change of speaker, or every minute or two, so you can find the audio behind any line in seconds.
• Number the lines (Word can do this automatically), so your qualitative coding and your quotes can cite an exact location.
• Mark what you couldn’t hear as [inaudible] or [unclear] with a timestamp, rather than guessing. The Minnesota Historical Society’s transcribing guidelines use the same convention, and it tells anyone checking your work exactly where to listen.

4. Check each transcript against the audio

This is the step people skip, and it’s the one I’d never let a student skip, because it’s what makes a transcript trustworthy. Play the whole recording while reading the transcript, and fix every word that doesn’t match. Do it soon after the interview, while you still remember what was said and how. If a transcription service typed it by hand, with people rather than software, you can focus your checking on the passages marked [inaudible], plus a spot check of any quote you plan to use in your findings.

It’s fallible even when you do it yourself. You’re tired, you know what your participant meant, and your brain fills in what it expects to hear. Slow down on the passages you’re most likely to quote, on technical terms, and on any speaker with an accent the software struggled with. One trick our coaches suggest for automated drafts: run the same audio through two different tools and compare them. The lines where they disagree are exactly where to listen hardest.

5. De-identify a copy, never the original

First, duplicate the checked transcript. The original stays untouched and stored securely, because it’s the evidence that your analysis reflects what was actually said. Every change from here happens on the copy.

On the copy, replace names with pseudonyms or codes (P1, P2), and remove or generalize anything that could identify someone: employers, towns, job titles, the name of the program they’re in. Keep a separate, secured key if your protocol allows one.

Pro tip: Strip the copy down to the questions and answers, cutting the greetings and small talk. It’s tidier to code, and it’s one less place for an identifying detail to hide.

Here’s what de-identification looks like on real speech. In another episode of our podcast, a dissertation chair credits her own chair with the advice she now gives her students, and names him, her university and herself along the way:

Notice that three details had to go: a named person, a named institution and the participant’s own name inside a quote, which is the easiest one to miss. The fillers and the false start went too, because this copy is for analysis. When you de-identify your own transcripts, search each one for names, places and organizations before you trust it, and keep the original exactly as it was.

6. Store it where your protocol says

File every recording and transcript where your data management plan says it’ll be, with the access controls it promised, and delete any stray copies:

• Files in your downloads folder
• Email attachments, sent and received
• The transcription tool’s or service provider’s copy

Then write the whole process into your methodology chapter: the device, the transcription route, the style, how you checked, how you de-identified and where the files live. Examiners ask about this, and a clear account is part of what makes a qualitative study credible.

Transcripts are only part of what they check in an interview-based study. Here’s a short walkthrough from our channel on what examiners expect to see:

An interview transcript example

Here’s the milestones excerpt from earlier, formatted as a research transcript with a header, speaker labels and timestamps. It’s a public podcast conversation between our host and a guest, not a research interview, so nothing is hidden. In a real study, the header and the text would carry pseudonyms only:

Take a moment to look at what the format buys you. The timestamp at 02:09 takes you straight to the audio behind the answer, the speaker labels carry no names, and “[laughter]” records something the words alone would lose. In your own transcripts, copy this shape and add line numbers when you open the file for coding.

---

Can AI transcribe your interviews?

Yes, as a first draft, where your ethics approval allows it. Automated transcription is fast and often good on clear audio, and our coaches often suggest it to students with a big stack of interviews. But you still have to manually review every transcript, word by word. And this requires a significant time investment.

When we produced the milestones excerpt, we transcribed it twice with two different automated tools and compared them. They disagreed in six places. One tool heard “far away” for “fire away”. More worrying, where the speaker started a word and abandoned it (“my dissert- or, you know”), one tool wrote “my district”, a real word that nobody said.

On the hippo excerpt in step 5, the same tool went the other way, finishing a half-word the speaker abandoned into “dissertation”. Both read smoothly, so nothing flags them, and in a research transcript they would sit there looking like data.

That’s the pattern to expect. Automated tools struggle most with accents, soft or overlapping speech and technical vocabulary, and they don’t mark their guesses, so the errors land in fluent-looking sentences. Used with the step-4 check, AI can save you some time. Used without it, you’re analyzing errors you can’t see.

I’d also keep general-purpose chatbots, such as ChatGPT, Gemini or Claude, out of it. A chatbot is the wrong tool for the job: it can tidy or summarize speech it was only asked to transcribe, and uploading participant audio to one is exactly the kind of data handling your protocol has to allow first.

---

How long does transcription take?

By hand, a long time. The Minnesota Historical Society’s guidelines give the industry standard as about one hour to transcribe fifteen minutes of audio, so budget roughly four hours of work per recorded hour, more for poor audio, several speakers or full verbatim notation.

An automated draft is quicker, but not free. Checking it means listening to the entire recording and correcting as you go, so budget more than the recording’s own length: for 15 interviews, 20 hours or more.

A transcript typed by a person needs far less: budget an hour or two to review anything marked [inaudible] and decide whether you can confidently fill it in, then spot-check the quotes you plan to use. Fifteen hour-long interviews typed by hand come to roughly sixty hours of work, which is why transcription deserves its own line in your project plan.

---

Where students often go wrong

These are the mistakes our coaches see most often, and every one of them is avoidable:

• Editing the only copy. Anonymizing or tidying the original leaves you with no raw data to verify against. Duplicate first, always.
• Trusting the automated draft. A transcript that nobody listened back to is a draft, however clean it looks.
• Assuming the recording will transcribe itself. In several meeting platforms, recording and transcription are separate settings. Test both before your first interview.
• Correcting a participant’s words. Fixing grammar or finishing someone’s sentence changes your data. Intelligent verbatim removes fillers; it doesn’t improve speech.
• Fearing the transcripts must be perfect before you start. They need to be accurate, not flawless. Mark genuinely inaudible passages and move on, rather than listening to the same four seconds twenty times.

Most of these cost a little time now or a great deal later, usually when an examiner asks how you checked your transcripts and the honest answer is “I didn’t”.

Still have questions?

Can ChatGPT transcribe an interview?

A general-purpose chatbot isn’t a reliable transcription tool, because it can tidy, shorten or summarize speech when you only asked for the words. Uploading participant audio to one is also a data-handling decision your ethics approval has to cover. If you use automated transcription at all, use a dedicated tool your protocol permits, and check its output against the full recording.

Do I have to transcribe every interview in full?

For most interview-based dissertations, yes, because thematic and similar analyses work from complete transcripts. Some approaches transcribe selectively, coding only the passages relevant to the research questions, but that’s a methodological choice you need to justify and agree with your advisor. If your methodology chapter promises full transcripts, transcribe them in full.

Can I correct what a participant said?

Not in the transcript itself. Fixing grammar or rewording changes your data; intelligent verbatim only removes fillers, repetitions and false starts. When you quote a participant in your findings chapter, you can shorten a quote with an ellipsis or add a clarifying word in square brackets, following your citation style.

Do transcripts go in the dissertation appendix?

It depends on your institution and your ethics approval. Some programs expect full transcripts in an appendix, others want a sample or none at all, often because full transcripts are harder to keep anonymous. Check your handbook, and if it’s silent, ask your advisor before you format an appendix.

Should I send transcripts back to participants?

Only if your research design or ethics approval includes it. Returning a transcript to its participant, sometimes called member checking or transcript review, lets them confirm accuracy and flag anything they regret sharing. It’s worth considering where interviews touch commercially or personally sensitive topics, but it adds time and needs to be planned in advance.

How long does it take to transcribe an interview?

The generally accepted rule of thumb is that transcribing by hand takes about one hour per fifteen minutes of audio, so roughly four hours per recorded hour. Poor audio, several speakers or full verbatim notation push that higher. Checking an automated draft is faster, but you still need to listen to the whole recording and correct as you go. A transcript from a human service mainly needs its [inaudible] passages reviewed.

Can’t find your answer here? Ask a Grad Coach directly – the initial chat is free.

Speak with a friendly coach →