Transcription is the first step of analysis, not a clerical task
Before it is a document, a transcript is a set of decisions. Bucholtz shows that transcribing involves both interpretive decisions (what gets represented) and representational decisions (how it is rendered on the page), and that an objective transcription is not possible (Bucholtz 2000). What you leave in and what you smooth out will steer the reading that follows.
Julia Bailey makes the same point for applied health research: putting speech into written form is an interpretive process and the first step in analysing data (Bailey 2008). Treat the transcript as analysis in progress. The person who transcribes is already coding, informally, by choosing where a sentence ends and whether a laugh belongs on the page.
That does not mean typing every word yourself. Manual transcription can run up to six hours for a single hour of audio (Haberl et al. 2023, citing Bell et al. 2018), which is time taken away from analysis. A practical route is the AI first pass, then a careful clean that experienced interviewers use: let a tool produce a draft, then read it against the audio and make the interpretive calls yourself.
Naturalized and denaturalized transcripts answer different questions
The verbatim question has a standard vocabulary. Oliver, Serovich and Mason distinguish naturalism, where every utterance is captured in as much detail as possible, from denaturalism, where grammar is corrected and interview noise like stutters and pauses is removed (Oliver, Serovich & Mason 2005). Neither is more correct; they suit different analyses.
Match the style to your method. Conversation analysis and discourse work need a naturalized transcript, because pauses, overlaps and false starts are the data. Thematic or content analysis usually reads better denaturalized, where the focus is what participants mean across cases. Decide before you edit, since retrofitting one style onto the other means re-listening to everything.
Whatever you pick, apply it to every transcript in the study so cases stay comparable. If you are writing up a thesis, a dissertation methods chapter has to justify that choice and show an examiner how much transcription detail you kept.
Which transcription conventions should you document?
Rigor here is mostly about what you write down. Poland argued that transcription quality is an aspect of trustworthiness, and that methods sections should report the steps taken to ensure audiotape quality, the directions given to transcribers, and an assessment of the transcript's trustworthiness (Poland 1995). A reviewer cannot judge a transcript when the making of it is invisible.
For any project with more than one transcriber, write the rules down. McLellan, MacQueen and Neidig note there are no universal transcription formats adequate for all qualitative data, so teams need systematic, documented procedures for preparing text (McLellan, MacQueen & Neidig 2003). A one-page style sheet settles the recurring questions.
Cover the decisions that come up on every line: how you mark pauses and overlaps, whether you keep fillers, how you spell non-standard speech, what a laugh or an [inaudible] gap looks like, and how you label speakers. Store that sheet with the data so a second coder, or your future self, transcribes the next file the same way.
Anonymization and consent set the boundaries of what you keep
Most qualitative recordings are human-subjects data. Under the Common Rule, a human subject is a living individual an investigator studies, and identifiable private information is information for which the subject's identity may readily be ascertained (45 CFR 46.102). A transcript full of names, employers and places is exactly that kind of information. Institutions also differ on which transcription tools they allow: in a survey of 130 universities' public guidance, most tool-level guidance came from IT and security pages rather than ethics offices, so check both before you upload.
If your data touch health information, the HIPAA Safe Harbor method lists 18 identifiers that must be removed, from names through any other unique identifying code (45 CFR 164.514). Even outside HIPAA, that list is a useful checklist for what to strip from a transcript before it circulates.
For EU participants, remember that pseudonymization is a weaker safeguard than anonymization. GDPR Recital 26 states that pseudonymized data still counts as personal data, and only truly anonymized data falls outside the rules (GDPR Recital 26). In practice, replace names with role labels as you clean the draft, keep an access-controlled master with the real identities, and use a tool that does not train on your files. Pepys never trains on your audio or transcripts and requires an account before provider-backed transcription starts.
How do you check an AI transcript against the recording?
Treat the AI output as a first draft and check it against the audio before you code or quote anything. Listen while you read, and correct names, terms, numbers, negations and who is speaking. An AI draft is a working document, not a certified or court-official transcript.
You do not need to replay every minute to catch most problems. A practical heuristic is to check several short stretches from the start, middle and end, plus every passage you will quote. If errors cluster, because of crosstalk, an accent, a noisy room or a specialist vocabulary, check the whole recording. This is a working method, not a statistical guarantee.
If you want a number, word error rate is the usual measure: count the substitutions, deletions and insertions against a reference you have corrected by hand, then divide by the number of words in that reference. As an illustration only, 3 errors in a 100-word sample is a 3% word error rate. Measure it on your own audio, because accuracy depends heavily on the recording. AI transcription accuracy statistics shows how published figures are measured.
| Check | Why it goes wrong | How to check |
|---|---|---|
| Names and technical terms | Rare words and proper nouns are the most common misheard items | Search the transcript for each participant, place and term, and fix them in one pass |
| Speaker changes | Interruptions and short replies are easy to attribute to the wrong voice | Play every turn change, then rename Speaker 1 and 2 to your participant codes |
| Numbers and negations | A dropped 'not' or a wrong figure reverses the meaning | Play back every sentence that contains a number or a negative |
| Overlapping speech | Two people talking at once usually becomes one person's words | Mark it and note who spoke, or listen again at the timestamp |
| Quotes you will publish | A quote has to match the audio exactly | Play each quoted passage and keep its timestamp |
Preparing a qualitative research transcription for coding
Coding starts from a clean, speaker-attributed file. Get the interview or focus group down to a first-pass draft with timestamps, correct whatever the tool misheard against the audio, then decide on your import format before you leave your editor. For the end-to-end workflow in one place, see research interview transcription, and for group sessions, how to transcribe a focus group.
Your CAQDAS software is particular about formats. NVivo imports plain text, CSV/TSV, rich text or Word, with an optional speaker-name column. ATLAS.ti accepts .doc, .docx, .rtf, .odt and .txt. Dedoose supports .doc, .docx, .rtf, .txt and .pdf. A DOCX export is the safe common denominator across all three, so a single file imports cleanly whichever package your team uses.
Get those four decisions right, style, conventions, anonymization and export format, and the transcript will hold up when a reviewer asks how it was made.


