Transcription looks like a clerical step between collecting data and analysing it, but the choices you make here shape what you can do later. How verbatim is verbatim? Who is "P3"? Where do the timestamps go? This guide walks through the decisions to make before you start, how manual and automatic transcription compare, and how to prepare transcripts for your analysis software and your methods chapter.
Decide how verbatim you need to be
There is no single correct level of detail. It follows from how you will analyse the data:
- Detailed transcription records how things were said: pause lengths, overlaps, stress, intonation, in-breaths. Conversation analysis uses systems such as the conventions described by Jefferson (2004). It is slow to produce and only worth it if those features are what you analyse.
- Full verbatim records every word as spoken, including false starts and fillers, without phonetic detail. Braun and Clarke (2006) ask for this kind of complete orthographic transcript for thematic analysis, with non-verbal sounds noted where they matter.
- Clean verbatim keeps the words and sentence structure but drops fillers, stutters and repetitions. It reads well and is common for content-focused analysis, such as many applied interview studies.
A summary of what someone said is not a transcript at all, and most qualitative methods do not accept it as data. Agree the level with your supervisor and write it down before the first interview is transcribed.
Conventions to fix before you start
- Speaker labels. For example I for the interviewer and P1, P2 for participants. Never real names.
- Timestamps. At least at the end of every speaker turn, so any passage can be traced back to the recording.
- Unclear passages. A marker such as [inaudible] with a timestamp, and a separate marker for words you are unsure about.
- Non-verbal events. Laughter, long pauses or interruptions in square brackets, if your analysis needs them.
- Identifying details. How names, places and organisations will be replaced, for example [Person 1] or [Hospital].
Manual or automatic?
Manual transcription takes several times the length of the recording; Dresing and Pehl (2018) put it at five to ten times. For twelve 45-minute interviews that is somewhere between 45 and 90 hours.
Automatic transcription removes the typing but not the checking. The output reads fluently and is still wrong in predictable places: names and technical terms, numbers, short words like "not" that reverse a meaning, speaker attribution for short interjections, and crosstalk, where usually only one voice is transcribed. In unclear passages the software tends to write something plausible rather than leave a gap, which is harder to spot than a visible blank.
Checking an automatic transcript
A fixed order makes the correction pass much quicker:
- Fix speaker labels first, merging any person split into two speakers.
- Correct recurring errors, such as a misheard term, across the whole transcript at once.
- Listen to the full recording once while reading along, and stop wherever something sounds off.
- Listen again to every passage you plan to quote verbatim.
Correcting does not mean polishing. If your conventions keep participants' sentence structure, an awkward sentence stays awkward.
Pseudonymise before analysis
Replace names and other identifiers before transcripts are shared or coded, and keep the key that links pseudonyms to real names separately from the data. Remember that removing names is not the same as anonymising: a job title in a small organisation can identify someone just as well. Under the GDPR, pseudonymised data is still personal data.
Formats for your analysis software
Most qualitative data analysis packages import Word documents. If you want to click from the text into the recording, timestamps have to be in a format the package recognises. MAXQDA, for example, reads several formats, including #hh:mm:ss-x# as used by f4, and recommends a timestamp at least at the end of each paragraph; it also imports VTT and SRT subtitle files. Check your package's manual before you export twenty transcripts in the wrong format.
A sentence for your methods chapter
Interviews were audio-recorded with participants' consent and transcribed in clean verbatim. Automatic transcripts were checked in full against the recordings and corrected; inaudible passages were marked. Participants are identified by codes (P1 to P12), and names, places and organisations were pseudonymised before analysis.
What software can and cannot do here
Themera transcribes English and German recordings with separated speakers and a timestamp for every passage. In the editor you rename and merge speakers, click a timestamp to hear the passage, fix recurring errors with find and replace, and pseudonymise names, places and organisations before analysis. Exports include Word with line numbers, MAXQDA-style timestamps, SRT, VTT and CSV. Data is stored and processed in the EU (Mistral AI, Paris) and not used for training. More on the transcription page.
What no software decides for you is the level of detail your method needs, and whether a misheard word changes the meaning of a quote. Once your transcripts are ready, the guide on how to code qualitative data covers the next step, and the focus group analysis guide covers group data.
Conclusion
Choose the level of detail from your method, fix the conventions before the first transcript, and treat an automatic transcript as a draft that you check against the recording. That combination saves most of the time transcription used to take without weakening your data.
See a sample analysis or try it with your own recording: three analyses free, no card needed, transcription included.