Interview Transcription Guide: Research, Journalism and HR

Interview transcription turns a recorded conversation into a written record you can quote, code, search and archive. The right way to do it depends on who you are: a qualitative researcher needs a consistent, analysable text; a journalist needs accurate quotes fast; an HR team needs a defensible record of what was said. This guide covers the decisions that matter across all three, from transcription style and speaker labels to anonymisation, consent and the choice between manual, automatic and hybrid workflows.

Who transcribes interviews, and what each group needs

The common thread: the transcript is evidence. It should be clear what was said, by whom, and what you changed.

Verbatim, clean verbatim and intelligent verbatim

These three terms describe how much of the raw speech survives into the text.

Verbatim (full verbatim) keeps everything: fillers ("um", "like"), false starts, repetitions, stutters, non-verbal sounds ([laughs], [sighs]), pauses and overlapping speech. Use it when how something was said matters: conversation analysis, discourse analysis, linguistic research, legal contexts.

Clean verbatim keeps every meaningful word but removes fillers, false starts and repetitions, and tidies grammar only where it does not change meaning. This is the default for most qualitative research, journalism and HR, because it reads naturally while remaining faithful.

Intelligent verbatim (or edited transcription) goes further, rephrasing for readability and dropping tangents. It produces a pleasant document but is no longer a faithful record, so it is unsuitable for analysis or anything that might be challenged.

The same sentence in three styles:

• Verbatim: "So, um, I think — I think the, the main problem was, you know, the onboarding, [laughs] it was just confusing."
• Clean verbatim: "I think the main problem was the onboarding. It was just confusing."
• Intelligent verbatim: "The main problem was the confusing onboarding."

Whatever you choose, write it down in a transcription protocol and apply it to every interview in the study. Mixed styles make comparisons unreliable. For research, your methodology section should state the style; many journals and supervisors expect it.

Before the interview: consent and recording

The transcript is only as good as the recording, and only as usable as the consent allows.

Consent

• Tell participants that the interview will be recorded and transcribed, who will do the transcription (you, a colleague, a service), where the files will be stored and for how long.
• If you use a cloud transcription service, name it in the consent form or information sheet; under GDPR this is part of informing participants about processing.
• Record verbal consent at the start of the audio as well as collecting the signed form. It travels with the file.
• For journalists, the rules differ by country and outlet; the minimum is that the interviewee knows they are being recorded and whether the conversation is on the record.
• For HR, inform the candidate or employee in writing before the meeting, and check local employment law and works council agreements; some jurisdictions restrict recording of interviews entirely.

Recording setup

• One microphone per speaker is ideal; a single phone on the table between two people is acceptable; a laptop across the room is not.
• For video calls, use the platform's recording with separate audio tracks where available. See our meeting transcription guide for Zoom, Teams and Meet specifics.
• Record a backup on a second device. Recordings fail more often than people expect.
• Start with a short slate: date, interview ID, interviewer initials. It makes the transcript self-identifying.
• Name files consistently (P07_2026-03-12_interview.m4a) and keep the key that maps IDs to participants in a separate, access-controlled place.

Speaker labels: what automatic diarization can and cannot do

Modern speech-to-text services identify speakers automatically (diarization) and label turns as Speaker 1, Speaker 2 and so on. In practice:

• Two speakers with distinct voices on a decent recording: usually reliable, with occasional mis-attributed short turns.
• Focus groups of five to eight: expect more errors, especially in cross-talk and with similar voices; plan for manual correction.
• Phone-quality audio: diarization degrades faster than word accuracy does.

The fix is a quick pass in an editor that lets you rename and reassign speakers while listening. In Audiotexta, for example, you rename "Speaker 1" to "Interviewer" once and the label updates throughout; clicking a line plays that moment so you can verify who spoke. Budget this pass into your timeline; it is the single most valuable correction step for interviews.

For research, decide early whether speakers are labelled by role (Interviewer / Participant), by ID (I / P07) or by pseudonym. Avoid real names in the transcript file itself.

Timestamps, inaudibles and conventions

A short set of conventions keeps transcripts consistent across a project. A reasonable default:

• Timestamps at every speaker change or every minute, in [hh:mm:ss] form. They let you return to the audio when a coded passage looks odd.
• [inaudible 00:14:22] for words you cannot make out, with the time, rather than a guess.
• [unclear: "scaling"?] when you have a best guess but want to flag uncertainty.
• [crosstalk] where speech overlaps and cannot be separated.
• [pause] or (3s) for pauses that matter to the analysis; omit in clean verbatim unless relevant.
• [laughs], [phone rings] for non-verbal events that affect meaning.
• Redactions marked explicitly, e.g. [name redacted] or [company], never silently deleted.
• [sic] for errors of fact or grammar you are deliberately preserving in a quote.

Keep the list in your protocol document and share it with anyone who touches the transcripts.

Anonymisation and pseudonymisation

Anonymising an interview means removing or replacing anything that could identify the participant, directly or in combination. It is harder than deleting names.

• Direct identifiers: names, employers, email addresses, phone numbers, exact job titles, specific locations.
• Indirect identifiers: unusual roles ("the only female engineer on the team"), rare events, dates combined with places, distinctive phrasing in small communities.
• Third parties: people the participant mentions also have privacy rights. Replace their names too.

Pseudonymisation replaces identifiers with codes or fictional names while a separate key allows re-identification by authorised staff. Under GDPR, pseudonymised data is still personal data; truly anonymised data is not, but true anonymisation of rich interview text is rare. Treat transcripts as personal data unless you have a documented reason not to.

A practical workflow: transcribe first, then anonymise in the editor with find-and-replace for known identifiers, then read through for indirect ones. Keep the original recording and the anonymised transcript under different access controls. Automatic entity-redaction features in some tools (names, emails, numbers) are a useful first pass but miss indirect identifiers, so a human read remains necessary.

For journalists the calculus is different: sources may be on the record, on background or anonymous, and the transcript should mark which. Store recordings of confidential sources with the same care as the notes.

Manual, automatic or hybrid?

The hybrid approach is now the pragmatic default. Automatic transcription produces a timestamped, speaker-labelled draft; you correct it while listening, applying your conventions. The correction pass is also where you build familiarity with the data, which qualitative researchers often cite as a benefit of transcribing at all.

On tools: Whisper run locally is free and keeps audio on your machine, at the cost of setup and no built-in editor (see our Whisper vs cloud APIs comparison); Otter, TurboScribe and similar services offer editors and summaries; Audiotexta adds speaker labels, an editor with playback highlighting, and export to DOCX, TXT and PDF, which cover the formats that NVivo, ATLAS.ti and MAXQDA import. See how to transcribe audio to text for a general walk-through and transcription accuracy tips for getting a cleaner first draft.

Role-specific workflows

Qualitative researchers

1. Write the transcription protocol (style, conventions, labels) before the first interview.
2. Transcribe each interview within a few days while the context is fresh.
3. Correct and anonymise in the editor; export DOCX with timestamps and speaker labels.
4. Import into your CAQDAS tool; keep the audio linked by file name for member checks.
5. Store the recording-to-ID key separately and record the retention date from your ethics approval.

Journalists

1. Transcribe immediately after the call; an automatic transcript catches quotes you half-remember.
2. Use timestamps to verify any quote you will print against the audio.
3. Mark on-record and off-record sections in the transcript.
4. Keep the recording and transcript for the period your outlet's policy requires.

HR teams

1. Confirm the legal basis and notify the interviewee in writing before recording.
2. Use a structured question list so transcripts are comparable between candidates.
3. Label speakers by role; keep commentary out of the transcript itself.
4. Apply a fixed retention period and delete recordings and transcripts when it ends.
5. Keep transcripts out of general shared drives; access should match the sensitivity.

Data protection notes for EU interviews

Interview recordings are personal data, and often special-category data when health, politics, religion or ethnicity come up. When using any transcription service, check where audio is processed, whether it is used for model training, how deletion works, and whether a data processing agreement is available. Our GDPR and transcription guide lists the questions to ask; Audiotexta's answers are on the privacy page.

A good interview transcript is boring in the best sense: consistent, labelled, timestamped and honest about what could not be heard. Get those four things right and every downstream step, from coding to quoting, becomes easier.