Qualitative interview transcription: Theses, focus groups and oral history

Three passes for a thesis interview

1. The machine pass. Give the recording as a file and get the first text. With a microphone it can also be written during the interview, but keep your attention on the participant.
2. The listening pass. Read the whole text along with the sound. Clicking a paragraph’s time takes the recording there. The More → Original text window shows the difference between what the engine wrote first and the text now, and Find deleted lists what you removed with its time. So you can say in your methods what you changed and how.
3. The coding pass. Turn names into codes, generalise details that identify a person, and export the file.

Do not skip the listening pass. Braun and Clarke’s guide to thematic analysis counts transcription as part of getting to know your data. When a machine does the typing, listening is how you get that back.

Research interview transcript format

A format that works in most analysis tools:

Interviewer: How did the first year go?

P3: It was hard, at the start. [pause] Nobody told us about the new rules.

One paragraph per turn, a code in place of a name, and marks in square brackets. LiveType exports speaker names at the start of a paragraph whenever the speaker changes, so naming the labels gives you this layout. Keep a header with the interview code, date and interviewer in a separate sheet or at the top of the file.

Phenomenological research: Keep their words

In phenomenology the participant’s own words for their experience are the point, so correction stays minimal: Awkward order, unfinished thoughts and repetition remain. LiveType does not write hesitation sounds like “um” and does not mark pauses. If they matter to your analysis, add them by ear. The model also sometimes skips words like “you know” on its own.

Focus groups: Many speakers, overlapping voices

Focus group transcription is the hardest case. Put two recorders on the table and have everyone say their code at the start. Speaker separation is on by default: Replace the labels with codes like “P1” and “P2”. It may split one person into two numbers; give both the same code and they merge. If two people end up under one number, naming cannot separate them and those parts must be corrected by hand. In overlaps the model usually writes one voice, so fill in the other by ear.

Anonymising participants

Removing names alone is not enough: In a small organisation a job title or an event identifies someone too. Keep the key separate from the text, and encrypted.

One thing to watch in LiveType: Original text and Versions keep earlier states, so a name you changed in the text still shows there. Export the anonymised text to share it, and delete the transcript when you finish. Keep this in mind when sharing that transcript with a team or by link.

Ethics review: What to write

Your application and methods should answer: Where is the sound processed, where does the text sit, who can open it, when is it deleted, and is there a transfer abroad? With on-device transcription in LiveType, a statement like this is accurate:

Interviews were transcribed on the researcher’s computer by an open-source speech recognition model (Whisper) running in the browser; audio files did not leave the device. Texts were stored on the tool’s server during the study, corrected against the audio, and deleted after being anonymised and exported.

With cloud transcription this is not true: The sound goes to a provider, often abroad. Private transcription explains the difference. Speaker separation also makes a voice print on the device, which may count as biometric data: If your consent form does not cover it, turn it off in Settings. For Turkish law see KVKK and voice recordings, and check your institution’s rules.

Importing into analysis tools

DOCX, ODT, TXT and Markdown files from Export go into qualitative analysis programs as documents. Speaker names are in the text; paragraph times are not in these files.

If you need a time link between text and sound, download SRT or VTT. MAXQDA’s help lists these formats too, with steps for importing transcripts with timestamps. Subtitle files split paragraphs into short lines, so DOCX is more comfortable for reading and coding. For NVivo and ATLAS.ti see their own help pages. For the analysis stage, discourse and thematic analysis may help.

Oral history: The recording is the source

In oral history the record is the recording, and the text is one reading of it. Baylor University’s guide to transcribing advises changing the narrator’s word choice and way of speaking as little as possible, and reminds you that a transcript loses things. The Oral History Association’s archiving manual notes that voice recognition quality varies widely and that first drafts must be reviewed.

You need little equipment: A recorder, a second device as backup, a lapel microphone if you can, and a quiet room. In microphone sessions LiveType also keeps the sound on your device under More → Recordings, where you can download it. Browsers may delete such data when space runs low, so download and archive it the same day.

Elderly narrators, dialects and place names

The models were not measured on dialects and elderly speakers. In such recordings automatic text is only a draft, so plan to listen word by word.

• Keep the way they speak. The model may turn dialect into standard language; restore it from the sound if it is part of your study.
• Write the rule down. Decide how far you will reflect pronunciation and apply it to every interview.
• Ask the narrator about names. Old place names and surnames go wrong most often.

A consent and archive card

Fill in the same card for each interview and keep it with the text:

What transcription does not add

Writing an interview down is not coding, finding themes or interpreting: LiveType does none of those. Before resting findings on the text, compare one full interview with the sound and see the level of error for yourself. The free plan has 30 minutes in total; for all of a study’s interviews see the plans.