Interview Transcription for Analysis: Clean Verbatim Guide

Interview transcription is where most qualitative projects lose their first week, and the loss is rarely the typing. It is the export. Zoom, Microsoft Teams, Otter and Rev all produce text, but they produce it for meetings and captions, not for coding: a timestamp on every line, one answer split across twenty cues, “Speaker 1” where the transcript needs Participant, and every um left in. An analysis-ready interview transcript has four properties: one paragraph per speaker turn, one consistent label per speaker, no timestamps in the running text, and a written set of conventions that says exactly what was kept and what was removed.

This guide covers the three levels of verbatim, what breaks in each export format and how to fix it, a cleaning workflow, the sentence to put in your methods section, and when cleaning is the wrong thing to do.

What does an analysis-ready interview transcript look like?

An analysis-ready interview transcript is one a coder can read and segment without fighting the formatting: each speaker turn is a single paragraph, every speaker has one stable label, timestamps live in metadata rather than in the text, and the transcription conventions are written down (McLellan, MacQueen & Neidig, 2003). McLellan and colleagues, writing from multi-site public-health studies, concluded that no universal transcription format is adequate for every design or theoretical framework, but that a documented, consistent format is what lets a team organize and analyze text systematically.

Those four properties are structural, and they are the ones meeting software gets wrong. Accuracy of the words is a separate question, covered below. A perfectly accurate Zoom caption file is still not a transcript.

Clean verbatim vs full verbatim vs Jefferson: which level do you need?

Full verbatim keeps every utterance and sound, clean verbatim removes fillers, stutters and false starts but keeps every content word, and Jefferson notation adds overlap, timing and prosody for conversation analysis. The level of verbatim is a decision about what counts as data. Oliver, Serovich and Mason (2005) describe a continuum from naturalism, where every utterance and sound is captured as fully as possible, to denaturalism, where idiosyncratic speech features are removed and grammar is tidied so the content stands out. Neither is correct in the abstract; each privileges a different reading of the interview.

Level	Keeps	Removes	Use it for
Full (true) verbatim	Every word, filler, false start, repetition, sound event ([laughs], [pause])	Nothing	Conversation and discourse analysis; studies of fluency, hesitation, affect or interaction
Clean verbatim	Every content word, the speaker's grammar and word choice	Filled pauses (um, uh), stutters and false starts, immediate repetitions	Thematic, content and framework analysis; most interview studies
Readable / edited	Meaning	Fillers plus hedges set off by commas (you know, I mean), sound tags	Reports and quotations for lay readers; not a basis for coding
Jefferson notation	Everything, plus overlap, pauses in tenths of a second, intonation, stress, volume	Nothing	Conversation analysis

Clean verbatim is the default for analysis of what people said, because removing “um” changes no content word and makes the text far easier to segment. Halcomb and Davidson (2006) went further and questioned whether full verbatim transcription of every interview is necessary at all in some qualitative designs, proposing field notes, reflective journaling and repeated listening as an alternative for some designs. That is a defensible position for some projects. It is a poor fit for any study where the speech itself is the object: the Labov and Waletzky narrative structure scheme depends on clause-level ordering, and the fluency and stuttering measures and psychomotor speech signs are rated from exactly the disfluencies and pauses clean verbatim removes.

What breaks in Zoom, Teams, Otter and Rev exports?

Zoom and Teams caption files split one turn into many timed cues, an Otter transcript puts a speaker-and-timestamp header above every paragraph, and Rev-style files prefix each turn with a name and time; all of them need turns rebuilt before coding. Zoom transcript cleanup, in practice, means merging cues, stripping timestamps and relabelling speakers. The snippets below are synthetic, in the shape each tool produces.

Zoom and Teams caption files (WebVTT, SRT). Captions are timed cues of a few seconds each, so one answer arrives as a dozen blocks, each with its own timestamp line and, in Teams, a <v Name> voice tag:

WEBVTT

3
00:00:10.500 --> 00:00:14.960
Marcus Lee: Uh, it was, it was about three years ago. She had a fall,

4
00:00:14.960 --> 00:00:19.300
Marcus Lee: and and after that she couldn't really manage the stairs on her own.

What breaks: the turn is split mid-sentence, cue numbers and timestamps interleave the text, and if speaker identification was off in the meeting settings the file carries no names at all, so every cue lands on one speaker. Teams and Zoom also write system lines such as “started transcription” as if someone said them.

Otter transcript and Teams Word export. Otter puts a header line with the speaker and a timestamp above each paragraph, and splits a single turn into several paragraphs on pauses:

Speaker 2  0:07
Yeah, so, uh, I'd been having trouble sleeping for months. Like, months.

Speaker 2  0:19
I was I was a bit nervous about it, to be honest.

What breaks: consecutive paragraphs from the same speaker have to be merged, the headers have to go, and “Speaker 1” and “Speaker 2” have to be mapped to roles, consistently, across every interview in the study.

Rev-style turns. Rev and several services write the name and a timestamp in parentheses at the start of each turn, Name (00:01): text. The structure is close to right; the timestamps still have to be stripped from the text.

Copied Zoom transcript. Copying from the Zoom transcript panel puts the name and the time on separate lines above each utterance, which looks like three speakers where there are two.

Our transcript cleaner detects which of these formats it has been given, rebuilds the turns, strips headers and timestamps, and then applies each clean-verbatim rule as a visible list of removed ranges so you can see exactly what was taken out. It runs in the browser and nothing is uploaded.

How accurate is automatic interview transcription?

Automatic transcription is good enough to start from and not good enough to code from without a correction pass. In a pilot comparison on virtual psychiatric interviews, Seyedi and colleagues (2023) scored three automatic services and one human service against reference transcripts: the best automatic service, Amazon Transcribe, had a median word error rate of 8.9% (interquartile range 6.9% to 11.6%), the human service Rev 7.6% (5.4% to 11.35%), and Zoom-Otter and Whisper were significantly worse than Amazon. In that sample the gap between the best machine and the human service was 1.3 percentage points, and the errors concentrated in a few word categories (conversational, cognitive and function words) rather than being spread at random (Seyedi et al., 2023).

Bokhove and Downey (2018) reached the same practical conclusion from a proof-of-concept on an interview, a public hearing and a classroom recording: automated captions give a “good enough” first version whose persistent mismatches can be found and fixed in a review-and-edit pass, with considerable time and cost advantages over transcribing from scratch. The word “first” is doing the work in that sentence.

To know how good your own draft is, correct a sample of ten minutes against the audio and score the draft against the corrected version in the word error rate calculator. A WER in the single digits on clear two-person audio is typical; a WER above 20% usually means poor audio, overlapping speech or unfamiliar accents, and the correction pass will take longer than the estimate.

How do you clean an interview transcript, step by step?

Cleaning an interview transcript is a six-step job, and the order matters because de-identification has to happen before the text goes anywhere, and the verbatim level has to be fixed before anyone removes a word.

1. Decide the verbatim level and write the rules down. Which of fillers, stutters and false starts, comma-set hedges, and sound tags will be removed, and which kept. Apply the same rules to every transcript.
2. De-identify. Replace names, places, dates, employers and record numbers with consistent pseudonyms before upload to any cloud transcription or analysis tool. Our transcript de-identifier finds candidate identifiers and keeps the same pseudonym for the same person across interviews. The consent and data-handling side is covered in annotation data, consent and licensing.
3. Rebuild the turns. Merge split cues and consecutive same-speaker paragraphs into one paragraph per turn; remove cue numbers, timestamps, headers and system lines.
4. Label speakers by role. Map “Speaker 1” to Interviewer and “Speaker 2” to Participant (or a participant code), and check that the mapping is right in every file; automatic labels swap between recordings.
5. Correct the words against the audio. Listen to the whole recording with the draft in front of you, or at minimum every passage you intend to quote or code. Poland (1995) defines transcription error as a mismatch between transcript and recording, which is why this check belongs in the workflow rather than in the optional extras.
6. Apply the clean-verbatim rules, with a diff. Remove what the rules say and nothing else, and keep a view of what was removed. A rule that strips “like” as a filler will also strip “I like my job” if it is written carelessly, which is why the cleaner only removes hedges that are set off by commas.

What should the methods section say about interview transcription?

The methods section should state who transcribed, from what source, to what standard, how speakers were labelled, how the transcripts were checked, and how identifiers were handled. Poland’s complaint, thirty years on, still applies: attention to transcription quality had not become routine practice in qualitative research, and it should count as part of rigor (Poland, 1995). A paragraph like the one below, adapted to the study, is enough.

Interviews were audio-recorded and transcribed by the first author from the recordings, using an automatic draft that was corrected in full against the audio. Transcripts follow a clean-verbatim convention: filled pauses, stutters and false starts were removed and all content words retained; non-verbal events were noted in square brackets. Speakers were labelled by role (Interviewer, Participant). Names, places and dates were replaced with pseudonyms before analysis; the same pseudonym was used for the same referent across interviews. Transcription conventions were agreed before transcription began and are available on request.

If a transcription service was used, name it and say whether it was human or automatic; if the draft was automatic, say how it was corrected. Reporting checklists for qualitative studies ask for exactly these details, and the same transparency applies to the sample-size rationale, covered in how many interviews reach thematic saturation.

When should you not clean a transcript?

Do not clean a transcript when the speech itself is the data. Conversation analysis depends on overlaps, pauses and repair; discourse analysis often depends on hedges and fillers; clinical speech research rates the disfluencies, pauses and rate changes that clean verbatim removes. Oliver, Serovich and Mason (2005) argue that denaturalized transcripts discard features of speech that some analyses treat as data; hesitations and silences that signal discomfort with a topic are one example, which is why conversation and discourse analysts keep them.

Two other cases: quotation and secondary use. Published quotations drawn from a cleaned transcript should be flagged as cleaned, because readers assume verbatim. And a transcript archived for reuse is more valuable in full verbatim, since a later study can derive a clean copy but not the reverse. The practical rule is to store full verbatim, code from clean verbatim where the method allows, and keep both under the same file name.

For audio-based annotation, where the recording rather than the text is what gets coded, see the audio and speech annotation guide.

How long does interview transcription take, and what does it cost?

Typing an hour of interview takes a professional transcriptionist roughly 3 to 5 hours and an untrained researcher 6 to 10, depending on audio quality, the number of speakers and whether full verbatim is required; the industry benchmark most services quote is about 4 hours per audio hour for clear two-person audio. At a conversational rate near 150 words per minute, an hour of interview is about 9,000 words of transcript. Professional human transcription typically costs about $1.50 to $3.00 per audio minute, and automatic services about $0.10 to $0.40 per audio minute plus your correction time; correcting an automatic draft takes roughly 0.5 to 1.5 hours per audio hour, depending on audio quality and the number of speakers (industry figures, as compiled in our transcription time calculator). Automatic transcription produces the draft in minutes, after which correction against the audio is the cost, and it scales with the error rate.

Method	Hours per audio hour (clear, two speakers)	What you get
Type it yourself, beginner	6–10	Full control; slow; good immersion in the data
Professional transcriptionist	3–5	Clean or full verbatim to spec; typical rates per audio minute
Automatic draft + full correction	Draft in minutes, then a correction pass of about 0.5 to 1.5 h per audio hour	Fast; quality depends on audio and the correction effort

The transcription time calculator takes the recording length, the number of speakers, audio quality, accents, the verbatim level and the method, and returns hours, word count and cost as a band with the working shown. Budget the correction pass honestly; the draft is the cheap part.

What are the most common interview transcription mistakes?

The most common interview transcription mistakes are coding from an uncorrected automatic draft, applying different cleaning rules to different interviews, losing speaker identity in the merge, and uploading identifiable audio to a service outside the consent. Each is avoidable:

• Coding from the draft. A median word error rate near 9% means one word in eleven is wrong, and the wrong words cluster in conversational and function words (Seyedi et al., 2023). Correct against the audio first.
• Inconsistent rules. One transcript with fillers and one without will code differently. Fix the rules in writing and apply them to all.
• Speaker drift. “Speaker 1” is the interviewer in one file and the participant in the next. Check the mapping per file.
• Interleaved timestamps. Timestamps left in the running text break sentence segmentation in every analysis tool. Move them to metadata.
• Cleaning before deciding. Removing fillers from a study that later wants to look at hesitation. Store full verbatim; derive clean.
• Uploading identifiable recordings. De-identify, and check the consent form covers the service you are using.

The practical upshot: interview transcription is a methods decision wearing an admin costume. Choose the verbatim level for the analysis, write the rules down, correct the words against the audio, and fix the structure the export left behind. The tools we have linked do the structural part in a minute; the judgment is still yours. For how transcripts then move into coding, see our survey of clinical transcript annotation tools.

References

• McLellan, E., MacQueen, K. M., & Neidig, J. L. (2003). Beyond the qualitative interview: Data preparation and transcription. Field Methods, 15(1), 63–84. doi:10.1177/1525822X02239573
• Poland, B. D. (1995). Transcription quality as an aspect of rigor in qualitative research. Qualitative Inquiry, 1(3), 290–310. doi:10.1177/107780049500100302
• Oliver, D. G., Serovich, J. M., & Mason, T. L. (2005). Constraints and opportunities with interview transcription: Towards reflection in qualitative research. Social Forces, 84(2), 1273–1289. doi:10.1353/sof.2006.0023
• Halcomb, E. J., & Davidson, P. M. (2006). Is verbatim transcription of interview data always necessary? Applied Nursing Research, 19(1), 38–42. doi:10.1016/j.apnr.2005.06.001
• Bokhove, C., & Downey, C. (2018). Automated generation of ‘good enough’ transcripts as a first step to transcription of audio-recorded data. Methodological Innovations, 11(2). doi:10.1177/2059799118790743
• Seyedi, S., Griner, E., Corbin, L., Jiang, Z., Roberts, K., Iacobelli, L., Milloy, A., Boazak, M., Bahrami Rad, A., Abbasi, A., Cotes, R., & Clifford, G. D. (2023). Using HIPAA-compliant transcription services for virtual psychiatric interviews: Pilot comparison study. JMIR Mental Health, 10, e48517. doi:10.2196/48517