Transcribing Interviews for Research: Consent, Accuracy and Analysis
A rigorous workflow for transcribing research and user interviews with AI: informed consent, audio setup, verbatim standards, anonymisation and coding the transcripts.
Transcription in a research context is held to a different standard than transcription for meeting notes. A wrong word in a meeting summary is an inconvenience. A wrong word inside a quoted participant statement is a research integrity problem, and it is invisible, because AI transcripts always read fluently.
AI transcription is still worth using — it removes hours of mechanical labour per interview — but only inside a process that treats it as a draft.
Before the interview: consent that actually informs
Informed consent for recording covers more than "may I record this". Participants should understand:
- That you are recording, in what formats, and for how long you will keep them.
- That automatic transcription will be used, and that a third-party service will therefore process their audio.
- Who can access the audio and the transcript.
- How their data will be anonymised, and the limits of that anonymisation.
- How to withdraw, and what withdrawal actually removes.
Get it in writing beforehand. Then restate it in the first thirty seconds of the recording, briefly, so the consent lives inside the artefact and you never have to reconcile a form with a file.
If your consent form promises data residency in a particular region or promises that content is not used to train models, verify that your transcription tool honours both before the first interview. Discovering a mismatch afterwards is a disclosure event, not a configuration issue.
Audio setup, because it determines everything downstream
Every hour spent on audio saves several hours of correction.
Get close. A lapel or headset microphone on the participant, or a dedicated recorder within arm's reach. A laptop microphone across a table is the single most common cause of unusable research audio.
Record each speaker separately when you can. For remote interviews, per-participant tracks make overlap a non-issue and speaker attribution essentially perfect. This is the biggest structural improvement available to remote researchers and it is frequently free.
Control the room. Soft furnishings, closed windows, no air conditioning, no café. Steady background noise masks precisely the consonant detail that distinguishes similar words.
Record a backup. A second device running in parallel. Interview audio is usually unrepeatable, and asking a participant to redo a 90-minute session because of a file corruption is a conversation nobody wants to have.
Say a spoken slate. Study name, participant code, date, session number, in the first ten seconds. If filenames get scrambled — and they will — the file identifies itself.
The first pass, and what to do with it
Run the audio through automatic transcription. Expect a competent draft with characteristic weaknesses:
- Proper nouns, organisation names and technical terms guessed phonetically.
- Numbers occasionally wrong in ways that look plausible.
- Speaker labels degraded around overlapping speech.
- Filler and false starts either over-preserved or silently cleaned, depending on the tool.
Then decide your transcription convention before you start correcting, and apply it identically to every interview. Mixing conventions across a corpus quietly breaks cross-interview comparison.
Full verbatim keeps pauses, repetitions, false starts, interruptions and non-verbal sounds. Required for conversation analysis and discourse analysis, where how something was said carries the meaning.
Cleaned verbatim removes filler and stutters while preserving exact wording and meaning. Adequate and usual for thematic analysis and most applied user research.
Never paraphrase. A transcript that improves a participant's grammar has altered the data. If you need a readable version for a report, quote cleanly and note the convention.
Verification: the step that cannot be skipped
The dangerous property of AI transcription is fluency. A substituted word arrives inside a grammatical sentence and carries no confidence signal, so nothing in the text tells you to be suspicious. A human transcriber's uncertainty appears as "[inaudible]"; a model's uncertainty appears as a confident guess.
A proportionate verification policy:
- Every quote you will publish gets listened to. No exceptions. This is cheap when you have timestamps.
- Every number, date and name gets checked. These are the highest-error, highest-consequence categories.
- Any passage you code closely gets listened to at least once. Interpretation built on a misread sentence propagates through your entire analysis.
- Passages at speaker-label boundaries get verified for attribution. Overlap is where diarisation fails, and attributing a participant's statement to the interviewer is a serious error.
Mark uncertainty explicitly in the transcript rather than smoothing it over. Square brackets with "[unclear]" are more honest, and more useful to a reader, than a confident wrong word.
Anonymisation, done properly
Automated find-and-replace is not anonymisation. It misses nicknames, inflections, partial mentions and the participant referring to themselves in the third person.
Work through two layers.
Direct identifiers. Names of the participant, colleagues, family members. Replace with consistent pseudonyms or codes so that references remain traceable within the transcript — "P07 mentioned [Colleague A] twice" preserves the relationship without the identity.
Indirect identifiers. Employer, job title, city, team size, unusual conditions or events. Individually harmless, jointly re-identifying. "The only female engineer on a four-person team at a Tallinn logistics startup" identifies someone precisely even with every name removed. Generalise to the level your analysis actually requires.
Keep the mapping between codes and identities in a separate, access-controlled location, never in the transcript file. And write down your anonymisation rules so a second researcher applies them the same way.
Where AI helps with analysis, and where it must not
Once you have a verified corpus, AI is legitimately useful for the mechanical parts of qualitative work.
Cross-interview retrieval. "Every time a participant mentioned finding onboarding confusing" across forty interviews, in seconds. Manually this is hours, and manual scanning misses instances where the participant used different words than you were searching for.
Candidate theme surfacing. Asking what topics recur produces a starting list. Treat it as a prompt for your own thinking, not as a finding.
Quote retrieval. Finding the clearest expression of a theme you have already identified, with a timestamp so you can verify it.
Negative case search. Actively looking for participants who contradict an emerging theme. This is where AI is most valuable and least used, because it counteracts the confirmation bias that makes qualitative analysis fragile.
What it must not do is decide your themes. Thematic analysis is interpretive work grounded in your knowledge of the participants, the context and the research question. A model summarising forty transcripts produces the average of what was said, which is often precisely the opposite of the insight — the interesting finding is frequently the outlier that the averaging discards.
Every quote that reaches a report gets verified against audio. Every claim about frequency gets counted rather than estimated. Those two rules keep AI-assisted analysis defensible.
What automatic transcription costs you as a researcher
There is a loss that rarely gets discussed, and it is not accuracy.
Manual transcription is tedious, and it is also the most thorough reading of your data you will ever do. Transcribing an interview by hand means listening to every sentence two or three times, which is why researchers who transcribe their own material frequently report that their analysis began during transcription rather than after it. Themes surface because you were forced to sit with the audio.
Automating that step removes the tedium and the immersion together. A verified AI transcript is just as accurate, and you will know the material less well than if you had typed it.
This is a real trade-off, not a reason to go back to manual transcription — the time saving is too large to give up, particularly on studies with thirty or forty interviews. But it should be paid for deliberately rather than lost by default.
Two ways to pay it:
Listen once, fully, without stopping. Not to correct, not to code, just to hear the interview again as a whole. It takes the length of the recording and it restores most of the familiarity that transcription used to provide.
Write an immediate memo after each interview. Before reading anything, record or write two or three minutes on what struck you, what surprised you, and what you expected to hear but did not. Do it while the session is fresh. These memos are also data, and they capture your evolving interpretation in a way that a transcript never will.
Researchers who skip both tend to produce analyses that are technically sound and oddly thin — accurate about what was said, less confident about what it meant.
A checklist you can reuse
Before: written consent covering AI processing; tool data-handling verified against your promises; microphones tested; backup recorder charged; transcription convention chosen.
During: spoken slate; verbal consent restated; one person speaking at a time where possible; note timestamps of striking moments.
After: automatic first pass; verification of quotes, numbers, names and attribution; anonymisation in two layers; secure storage with the code mapping separated; retention window applied and documented.
Analysis: retrieval and negative-case search with AI; interpretation by you; every published quote checked against audio.
The discipline is the point. AI transcription makes research interviews dramatically less laborious, and it makes it dramatically easier to publish a sentence nobody ever said. The process above is what keeps you on the right side of that.
Related reading
- Speech-to-text accuracy guide — why fluent transcripts hide errors.
- How to transcribe meetings automatically — the lower-stakes version of this workflow.
- AI meeting notes for consultants — consent and records in client work.