Qualitative Research Transcription: Conventions, Ethics and Tools

Key Takeaways

Decide the level of detail before you transcribe anything: verbatim if your analysis is about how people spoke, intelligent verbatim or clean if it is about what they said. Write the convention down before the second interview. Check what your consent form and ethics approval say about where recordings may go, and ask any transcription service exactly who receives the audio. Automatic transcription gives you a first draft in minutes for a fraction of the cost of a human service; the correction pass against the audio is still yours. For NVivo, ATLAS.ti and MAXQDA, plan the import route before you choose the tool.

Nobody tells you the standard for qualitative research transcription because there isn’t one standard. There are several, and which one applies to you depends on what you plan to do with the transcripts once you have them. That decision is methodological, and it belongs before the tooling decision, not after it.

Get it the wrong way round and the cost is real. You transcribe six interviews cleanly, get into your analysis, discover you need to know where the participant hesitated, and now you’re going back through six recordings with different ears. So spend an hour on the question first.

Decide what the transcript is for before you transcribe a word

How much detail a transcript needs is a methodological decision, not a matter of care or thoroughness. It follows from your analytic approach.

If your analysis is concerned with what was said, a clean transcript is fine. You want the words, accurately, attributed to the right speaker. Filler words, false starts and short pauses can be tidied out without damaging anything, because none of them will ever appear in your coding. Most thematic and content-oriented work sits here, and a clean transcript is quicker to produce, quicker to read and quicker to code.

If your analysis is concerned with how something was said, the situation reverses. Pauses, overlaps, hesitations, the moment where two people talk over each other, the three seconds before an answer arrives, the repair mid-sentence: those are the data. A transcript that removes them has removed the thing you’re studying. This matters for automatic transcription specifically, because smoothing out exactly those features is what it does well. It gives you fluent, readable prose, and fluent readable prose is a lossy rendering of a hesitant, overlapping conversation.

There’s a temptation to hedge by transcribing everything to the finest level of detail, on the grounds that you can always ignore what you don’t need. Usually that’s the wrong move. Detailed transcription is slow, and slow work on twenty interviews eats months you don’t have. It also makes the transcripts harder to read, which makes coding slower. If you genuinely don’t know yet whether interaction matters to your analysis, that’s a conversation to have with your supervisor now, not a problem to solve by brute force. Whatever you decide, the recordings stay, and the recording is the actual data. The transcript is a representation of it.

Verbatim, intelligent verbatim and clean, side by side

The three labels get used loosely, so here is one short, made-up exchange written all three ways. It is an illustration, not data from a study.

Verbatim keeps everything the recorder heard, marked by your convention:

I:  So how did you, um, how did you find out?
P3: Well (.) I- I didn't, really. My sister [told me]
I:                                          [right]
P3: told me after the- after the meeting. (2.0) Yeah.

Intelligent verbatim keeps every meaningful word and the speaker turns, drops fillers and stutters, and keeps a pause only where it carries meaning:

I:  How did you find out?
P3: I didn't, really. My sister told me after the meeting. [pause] Yeah.

Clean (sometimes called edited) also smooths the grammar into readable sentences:

I:  How did you find out?
P3: I didn't. My sister told me after the meeting.

Look at what the clean version lost. The two-second pause before “Yeah” and the self-correction “I- I didn’t” might be the most interesting thing in the extract if you are studying how people talk about bad news. For a thematic study of how news spread through a family, the clean version loses nothing you will code.

Automatic transcription lands somewhere between intelligent verbatim and clean. It usually drops the “um” and the repeated “I”, writes no pause marks, and shows overlap as one turn following another. If you need verbatim, plan to add those features by hand in the correction pass.

Write your convention down before the second interview

Once you’ve decided the level of detail, fix the conventions and write them into a short document you keep beside you while you work. Changing them later means going back through every recording you’ve already done, and that’s the kind of task that quietly consumes a fortnight.

The things worth settling in advance:

• Whether filler words and false starts are kept, tidied, or kept only where they seem meaningful, which is the option most likely to produce inconsistency across interviews.
• How pauses are marked, and whether short and long pauses are distinguished from each other.
• How overlapping speech is shown, and what happens when you can’t tell who spoke first.
• How you mark passages you can’t make out, and whether you note the reason.
• How speakers are labelled, and how that labelling relates to your anonymisation scheme.
• Where timestamps go: at a regular interval, at each speaker change, or only against passages you expect to quote.

Timestamps deserve a moment of their own. They let a quotation in your thesis be traced back to its position in the recording, which is what makes the claim checkable by anyone who has access to your data, including an examiner. Adding them as you go costs almost nothing. Adding them afterwards, to a finished transcript, means listening to the whole recording again with a stopwatch. Whichever tool you use, get them in on the first pass.

A convention template you can copy

Fill in the right-hand column, keep it in your methods appendix, and give it to anyone who transcribes or corrects for you. The marks shown are common choices, not rules; change any of them, but write the change down.

The same template as plain text, for your notes:

TRANSCRIPTION CONVENTION: [project]          version: [date]
Detail level:        verbatim | intelligent verbatim | clean
Speakers:            I = interviewer; P1, P2 ... = participants
Fillers:             kept | removed | kept when meaningful
False starts:        kept, marked with a hyphen (I- I didn't)
Pauses:              (.) under 2 s; (3.0) = length in seconds
Overlap:             [ at start and ] at end of overlapping speech
Inaudible:           [inaudible hh:mm:ss]
Uncertain:           [word?]
Non-verbal:          [laughs] [sighs] [door closes]
Emphasis:            not marked | CAPITALS
Timestamps:          each speaker change | every 2 min | quotes only
Anonymisation:       at transcription; mapping kept separately at [location]
Other languages:     original, then {translation}
Checked against audio by: [name], on [date]

Consent and ethics approval decide where the audio can go

Your consent form is not paperwork that happened before the research started. It is a set of conditions on what you may now do with the recordings, and it usually covers three things: how the recording is stored, who is allowed to hear it, and when it gets destroyed.

Sending audio to a third-party transcription service is a disclosure. Someone or something outside your project now has a copy of a participant’s voice. That has to be compatible with what the participant agreed to. Sometimes it plainly is, because the form anticipated it. Sometimes it plainly isn’t, because the participant was told only the research team would hear the recording. Often the form is ambiguous, and that ambiguity is worth resolving with your supervisor or ethics committee before you upload rather than after.

Ethics approval frequently goes further and specifies where recordings may be stored, on what kind of device or institutional system, and for how long. The transcript inherits those conditions. So does anything derived from it: your coded files, your quotation extracts, the working document where you paste passages while drafting. Those all count as data, and they all belong inside whatever storage arrangement you were approved for. It’s easy to lose track of this once you’re deep in analysis and copies start multiplying across a laptop, a cloud drive and an email to yourself.

Practically, this narrows your options for tooling before you’ve compared any of them on quality or price. A tool that processes files on your own device is a different proposition from one that uploads them. If a provider cannot tell you clearly which companies receive the audio, treat that as an answer.

What an ethics board asks about a transcription service

Committees differ, but the questions repeat. Put the answers in your application, with the date you checked them, because providers change their partners.

1. Where is the audio stored, and in which country?
2. Which other companies receive it, and what do they receive? A transcription tool is often a front end for someone else’s speech recognition. Ask for the list of sub-processors.
3. Is the data used to train models? Ask about the provider and about each partner separately; one answer rarely covers both.
4. What does deleting a file remove, and are there backups that keep it longer?
5. Who can sign in, and is there an audit report, a data processing agreement or institutional sign-in?

Here are SozAI’s answers, as of 25 October 2026, so you can see what a complete set looks like:

• Recordings and transcripts are stored on our servers. Everything travels over HTTPS.
• Speech recognition is done by AssemblyAI, in the US, which receives the audio. If you use summaries, notes, quizzes or questions about a recording, the transcript text goes through OpenRouter to model providers (Anthropic, Google and xAI, with DeepSeek as a fallback). YouTube links go to the Google Gemini API; your own audio does not. Cloudflare carries the traffic and holds our encrypted database backups. The full list is on our sub-processor page.
• We do not sell recordings, and our own code never uses them to train models. We cannot make that promise on behalf of the partners above; their own terms apply.
• You can delete a recording or your whole account yourself, in the app or on the web. That removes the audio, transcript, notes, flashcards and quizzes from our servers. Encrypted database backups are kept for up to 12 months. Deleting in SozAI does not delete the copy held by the speech recognition partner.
• Sign-in is with Google, Apple or a one-time email code. We do not have a SOC 2 report, single sign-on or team admin roles, and SozAI is not HIPAA-compliant. More on security at SozAI.

If your approval says recordings must not leave institutional systems, none of the cloud tools named on this page qualify, ours included. Use your university’s own licensed service or transcribe offline.

Anonymisation happens in the transcript

The recording stays as it is. You can’t anonymise a voice by editing text, and you shouldn’t be trying to alter your primary data anyway. Anonymisation is something you do to the transcript.

That means replacing names, place names, employer names, job titles that identify one person, and the small circumstantial details that make someone recognisable to anyone who knows the setting. Do it consistently. If a colleague is [Colleague A] in one interview, she is [Colleague A] everywhere she appears, and the same convention applies across your whole set. Inconsistent pseudonymisation is worse than none, because it looks anonymised while still being traceable by someone who reads carefully.

If you keep a mapping between real identities and replacements, keep it separately from the transcripts, under the storage conditions your approval sets out, and be clear with yourself about why you’re keeping it and when it gets destroyed. In some projects there’s a good reason to hold it: you may need to re-contact participants, or link interviews across waves. In others there isn’t, and the safest mapping is the one that doesn’t exist. Deciding deliberately is the point.

One caution about doing this at speed. Find-and-replace catches the obvious names and misses the rest: the nickname used twice, the misspelling, the description of a building that only one place in the city matches. The anonymisation pass wants human eyes on the whole document.

A practical point for automatic tools: the speaker names you type into a tool are stored with the transcript. Label speakers as P1 and P2 from the start rather than renaming them later.

Focus groups: six to ten speakers

A focus group is the hardest recording you will transcribe. Automatic speaker separation works by grouping speech that sounds alike, and with eight people round a table that breaks down in predictable places: two participants with similar voices end up as one speaker, someone who only says “mm” or “yes” never gets a label of their own, and a burst of crosstalk is attributed to whoever spoke loudest.

In SozAI you cannot tell the system how many people were in the room; it decides from the audio. Expect to fix speaker labels by hand on every focus group, and budget for it. On a one-hour, eight-person group that is a real afternoon of work, mostly spent on short turns.

What helps, in order of effect:

• Open with a round in which each person says their participant number and one sentence. That gives you a sample of every voice at a known point, so you can match labels to people.
• Have the moderator name the speaker when it is unclear (“thanks, P4”) and keep a seating plan with the order of first contributions.
• Put one good microphone in the middle of the table rather than a phone at the moderator’s end. Distance is the single biggest source of errors.
• Run a second recorder at the other end of the table. When one file is unclear, the other often is not.

If your analysis depends on who said what in fast exchanges (conversation analysis of group dynamics, for example), an automatic first draft will save you less time than it does on one-to-one interviews. A human service that specialises in focus groups may be worth its price there.

Where automatic transcription actually helps

Automatic transcription produces a first draft. That’s a genuine saving, and refusing it on principle mostly means typing out several hours of speech by hand for no methodological gain. What it doesn’t produce is a finished transcript.

For interview data, the correction pass against the audio is part of the method, not an optional tidy-up at the end. You listen through with the draft in front of you, fix what’s wrong, restore whatever your convention requires that the machine dropped, insert or check the timestamps, and apply your anonymisation. It’s slower than reading and faster than typing. It’s also where you become familiar with the material, which is not a side benefit. Researchers who transcribe their own interviews arrive at analysis already knowing what’s in them, and the correction pass buys back a decent share of that familiarity.

With the SozAI app for iPhone, Android and Mac, or the web version, a recorded interview comes back with speaker labels and a timecode on every passage; tap one and the audio plays from that line, which is most of what a correction pass needs. Text can be corrected in the iPhone and Android app; the web and Mac versions show it but do not edit it yet. More on transcribing interviews in general. It is one option among several, and the storage and consent questions an ethics board asks should filter your shortlist before quality does.

AI or human academic transcription: cost, turnaround, accuracy

Most of the pages that rank for “academic transcription” are human services, so here is the comparison they leave out. Prices are as each provider published them on 25 October 2026.

Run the numbers on a typical study: twenty one-hour interviews, 1,200 minutes of audio. At the prices above that is $2,388 with Rev, $2,100 with Ditto and from $948 with TranscribeMe. On a SozAI subscription the whole set fits comfortably inside the hours, and the drafts are back the same afternoon.

The accuracy figures are not comparable, and it is worth knowing why. The human services state a guarantee on their own terms. Ours is a measurement with a published method: 200 short sentences read aloud in quiet rooms, from the open FLEURS dataset. Interviews are harder than that, and some languages are much harder (19.6% for Arabic, 50% for Kazakh in the same test). So the honest trade is this: automatic transcription costs far less and is near-instant, and in exchange you do the correction pass yourself. A human service charges you for doing most of that pass, and you still check the quotes.

When human transcription is the better buy: verbatim conventions with pause timing, poor audio, focus groups where attribution is the point, or a funder who will pay for it while your time is the scarce resource.

Transcription software for qualitative research

Six realistic options, with what each is good at. Details as published by each provider on 25 October 2026.

Our recommendation, given a choice. If you already have a MAXQDA or NVivo licence and your interviews are in a language they cover, use the built-in transcription: the transcript arrives in the project already linked to the audio, and you skip the import step entirely. If your interviews are in a language outside their lists, are recorded on a phone in the field, or you want summaries and translation alongside the transcript, SozAI is the better tool, with one import step. If your ethics approval rules out third-party AI processing, pay a human service that will sign your confidentiality terms, or transcribe it yourself.

Getting transcripts into NVivo, ATLAS.ti and MAXQDA

SozAI does not export Word (DOCX) files. Everything below works with what it does export: TXT, SRT, VTT and JSON (PDF too, but no analysis package wants a PDF transcript). The phone and Mac apps export all of these; the web version exports TXT, SRT and VTT. Rename your speakers to P1, P2 and so on before you export; those names are written into VTT, JSON and the app’s TXT. SRT never carries speaker names, so skip it for research.

ATLAS.ti

Add the audio file to your project first, then import the transcript to it. ATLAS.ti’s manual lists VTT, SRT, TXT, DOCX and RTF as supported transcript formats, and says it detects and codes speakers in VTT files that include speaker information. So export VTT from SozAI: you get the text synchronised with the audio and speaker codes in one step. (ATLAS.ti manual, checked 25 October 2026.)

MAXQDA

MAXQDA imports transcripts with timestamps under Import, Transcripts, and lists VTT among the formats it reads, alongside SRT captions and Word or RTF files. Choose VTT and point MAXQDA to the matching audio file when it asks. SRT also imports, but MAXQDA users report a fragmented layout, one caption line per paragraph, and you lose the speakers. (MAXQDA help, checked 25 October 2026.)

NVivo

NVivo 14 for Windows imports a transcript linked to audio from a plain text file laid out as a comma- or tab-separated table, or from Word or RTF. (NVivo help, checked 25 October 2026.) SozAI’s files are not in that layout, so you have two routes:

• Quick: export TXT from web.sozai.app with timecodes and speakers switched on. Each paragraph reads “[04:12] P3: text”. Import it into NVivo as an ordinary document. You can code it straight away; it just is not synchronised with the audio.
• Synchronised: export JSON from the app. Each segment has a start and end time in seconds, the text, and the speaker name. Open it in a spreadsheet (or ask a colleague with a few lines of script) to make three columns: timespan, speaker, text. Save as tab-separated text and import it as a transcript, mapping the columns when NVivo asks.

Whichever route you choose, do the correction pass before the import if you can. Corrections made in the phone app travel with the transcript into every later export, while edits made inside your analysis package stay there.

What the first draft won’t do for you

Expect the draft to be weakest exactly where interviews are hardest. Crosstalk confuses speaker separation, and two people talking at once tends to come out as one person saying something slightly incoherent. Strong accents, quiet participants, background noise from a cafe or a corridor, and field-specific vocabulary all reduce accuracy. Speaker labels are usually close but not reliable at the boundaries, particularly around short interjections.

And it will not give you prosody. No automatic transcript preserves the length of a pause or marks the point where the participant’s voice changed. If your analysis needs that, the machine draft is a scaffold for the words and nothing more, and you should budget your time accordingly. A colleague working on the same corpus with a content-oriented question will finish in a fraction of the time, and that’s not because they’re working less carefully. Their transcript is answering a different question.

Interviews that switch language are a special case. Automatic tools, SozAI included, usually detect one language per recording, so the passages in the second language come out worst. Mark them in your convention and check them first.

Budget more time than the arithmetic suggests, too. The correction pass is not one pass through the audio at normal speed; it’s stopping, rewinding, re-listening to the six seconds you can’t quite make out. Doing this on a full set of interviews is substantial work, and it is normal for it to be. If you want a sense of how the pieces fit together across a whole project, there’s a walkthrough of a thesis researcher’s workflow that follows one set of recordings from file to quotable transcript. If you are buying for a research group or a department, write to [email protected] about Team and Education plans; we invoice.

The last thing to hold onto is that the transcript is a representation, and every convention you chose is a decision about what to keep and what to lose. Write those decisions into your methods chapter. An examiner who knows why your transcripts look the way they do will read them as considered. One who doesn’t will wonder.