From Interview Recordings to Themes: Preparing Transcripts for Analysis

What thematic analysis asks of a transcript

Thematic analysis is a family of methods for identifying patterns of meaning across a set of qualitative data. The version most often taught in the social and health sciences is Braun and Clarke's reflexive approach, usually described in six phases: familiarizing yourself with the data, generating codes, constructing candidate themes, reviewing them, defining and naming them, and writing up. Other versions, such as codebook or framework approaches, use a fixed coding frame from earlier in the process. Whatever the variant, the transcript is the surface you work on, and its quality shapes what you can see.

Three properties matter most. The words must be right, especially the words that carry meaning for your research question. The layout must make speakers, turns and pauses visible enough to interpret. And every passage must be traceable to the original audio, because interpretation sometimes depends on tone, hesitation or laughter that the text alone does not carry.

Check the machine draft before you code anything

Many researchers now start from an automatic transcript. That saves typing, but a machine draft is not yet data you can analyze. Recognition models make errors that matter in qualitative work: a negation dropped ("I could care less" becoming "I couldn't care less"), a technical or local term replaced with a common word, a quiet aside missed, or, in long silences, a short phrase that nobody said. The article on Whisper-style hallucinations explains why those inventions happen and where to look for them.

Before coding, listen to each recording in full while reading the draft. This is slower than skimming, but it doubles as the first round of familiarization, so the time is not wasted. The guide to proofreading an AI transcript covers the mechanics of an efficient pass; for analysis, add three research-specific checks:

• Mark every place where you changed the meaning, not just the spelling. These are the passages where the machine would have misled you.
• Note emotional or paralinguistic features the draft leaves out, such as laughter, long pauses, crying or a change of voice, in square brackets.
• Add speaker labels yourself. Automatic drafts from most general-purpose tools, mydubly included, do not say who is speaking.

Decide early how much verbatim detail you need. Reflexive thematic analysis usually works well with a clean, readable transcript, but if hesitations or false starts are part of what you want to interpret, keep them. The comparison of verbatim and clean verbatim transcription sets out what each style keeps.

Lay transcripts out for coding

A consistent layout makes coding faster and makes the dataset easier to share with a co-coder. Agree on these conventions before you prepare the second transcript, not after the tenth:

• A header block with a participant pseudonym, interview date, interviewer initials, recording length and the file name of the audio.
• One paragraph per speaker turn, with a short label such as INT for interviewer and a pseudonym for the participant.
• A timestamp at the start of each turn, or at least every minute or two, in the same format throughout.
• Square brackets for anything you added: [laughs], [long pause], [name of employer removed].
• Line or paragraph numbering if your coding software or your supervisor expects it.

If you import transcripts into qualitative analysis software such as NVivo, ATLAS.ti, MAXQDA or a free alternative, test one file first. Some programs read timestamps in a particular format and link them to the audio; others treat them as plain text. Plain text or a word-processor file with simple paragraphs is the safest starting point.

Familiarization: reading with the audio

Familiarization is more than reading the transcripts once. It means immersing yourself in the data until you can recall who said what and how. Most researchers combine reading, listening and early note-taking.

A practical routine for each interview:

1. Read the corrected transcript straight through without coding, to get the shape of the conversation.
2. Listen again to passages that felt important or ambiguous, following the timestamps.
3. Write a short familiarization note: what stood out, what surprised you, what connects to other interviews.
4. Record initial ideas in a research journal, dated, so you can later show how your thinking developed.

Listening matters here because the same words can carry different meanings. "It was fine" said flatly after a long pause is not the same data as "it was fine" said briskly. A transcript linked by timestamps lets you return to that moment in seconds.

Coding: from passages to labels

Coding means attaching short labels to segments of data that relate to your research question. In reflexive thematic analysis, codes can be semantic (close to what the participant said) or latent (an interpretation of underlying meaning), and they evolve as you work. In codebook approaches, you apply a predefined frame and refine it.

A few habits that keep coding traceable:

• Code meaningful segments rather than single words. A segment might be one sentence or a whole answer.
• Give each code a working definition in a codebook or memo, even in reflexive work, so you remember why you created it.
• Keep the timestamp with each coded segment. If you export coded extracts later, the timestamp travels with the quote.
• Recode early transcripts once your codes have settled. The first two interviews are often coded more thinly than the rest.

Building, reviewing and naming themes

Themes are patterns of shared meaning organized around a central idea, not simply topic summaries such as "views on training." Moving from codes to themes usually involves clustering codes, sketching thematic maps, and then testing candidate themes against the data.

Reviewing has two levels. First, check that the coded extracts within each theme hang together. Second, return to the full transcripts and ask whether the theme still makes sense across the dataset, including passages you did not code. This is where clean, timestamped transcripts pay off: you can re-read and re-listen to the contexts around each extract rather than relying on the extracts alone.

Once themes are stable, write a short definition for each, explain its boundaries, and choose a name that conveys its central idea. Pick illustrative extracts that are accurate, attributed to a pseudonym and, where they are central to an argument, double-checked against the audio one last time.

Memos and the audit trail

An audit trail is the record that lets someone else follow how you got from raw recordings to your final themes. It is part of how qualitative researchers demonstrate rigour, and examiners and reviewers increasingly ask to see one.

Keep these pieces together in a project folder:

• The original audio files, unchanged, with consistent file names.
• Each transcript version: the machine draft, the corrected version, and any anonymized version for sharing.
• A transcription log noting who corrected each file, when, and which conventions were used.
• Codebook versions with dates, showing when codes were merged, split or dropped.
• Analytic memos and thematic maps, dated.
• A list of quotes used in the write-up with participant pseudonym, transcript name and timestamp.

When you cite or quote recordings in a publication, the guide to citing videos with timestamps shows common referencing patterns. Interview data usually stays confidential, so the timestamp lives in your audit trail rather than in the published reference.

Pitfalls and limits of machine transcripts in analysis

• Coding an uncorrected draft. Small recognition errors become codes, and codes become themes.
• Treating fluent text as faithful text. Modern recognition models produce tidy sentences and tend to drop fillers and hesitations, which may smooth away features you would want to interpret.
• Losing the link to audio. If you strip timestamps when anonymizing, keep a private mapping so you can still go back to the recording.
• Mixing conventions across transcripts, which makes co-coding and comparison harder.
• Assuming one listen is enough for difficult audio. Accents, crosstalk and noisy rooms can need a second pass on specific stretches.
• Forgetting the ethics of the tool. Sending participant audio to any external service should match what participants consented to and what your ethics board approved; check with them first.

Group discussions add their own difficulties, mainly overlapping speech and identifying who said what. The article on transcribing focus groups covers those; most of the analysis steps above still apply once the transcript is ready.

How mydubly fits a thematic analysis project

mydubly can produce the first draft. You choose an interview recording from your device, in common audio formats such as MP3, WAV or M4A or as a video file, up to two hours per file, and it returns a plain transcript, a timestamped transcript with [m:ss] labels, and SRT and VTT subtitle files. The spoken language is detected automatically, across 21 supported languages. The audio to text page shows the upload and output options.

What it does not do matters for analysis. Transcripts have no speaker labels, no notation for pauses or laughter, and no coding features. Recognition output tends toward clean text, so if you need true verbatim detail you will add it during correction. Whether you may use an external transcription service at all for your participants' audio is a question for your ethics board and data management plan, not for the tool; check before uploading. For the record, audio chunks and results are deleted within 30 minutes of a job finishing.

Cost is one credit per minute of audio, with a five-credit minimum, so a 60-minute interview costs 60 credits (6¢).

Next step

Take one interview, produce a draft, and time how long a full corrective listen takes, adding speaker labels, bracketed notes and timestamps as you go. That single file gives you a realistic estimate for the whole dataset and a template layout to reuse. Settle your conventions in writing, then move on to familiarization with the rest. If your recordings are interviews in general rather than research data, the interview transcription use case covers the simpler workflow.