An interview transcript is a written record of a spoken interview in which every turn is attributed to a named speaker. A usable one has a header with the date and participants, a body of labelled turns with timestamps, and square brackets for anything that was not speech.
That definition is easy. What guides rarely show you is a transcript that came out of a real recording, with the small mess a real recording produces. Almost all the spoken transcript examples you will find online were written to illustrate the rules rather than transcribed from audio, which is why they all look immaculate.
So this guide uses a real one. Below is a genuine oral history interview recorded on 5 January 1985 β Sally Stapp, then in her eightieth year, interviewed by the Sausalito Historical Society. The recording is in the public domain with no restrictions on use, so we can put it here in full. We uploaded the first five minutes to our own pipeline with speaker detection on, and what follows is what came back, unedited.
A real transcribed interview example
Play it first. Thirty-six seconds, and the transcript underneath is what our pipeline returned for this exact file β so every line can be checked against your own ears instead of taken on trust.
This is the raw output. No corrections, no tidying, exactly as the machine produced it:
```
[00:02] SPEAKER_01: Now it's working.
[00:04] SPEAKER_01: Now it's working.
[00:05] SPEAKER_01: This is Sally Stapp and it is January 5th, 1985.
[00:10] SPEAKER_00: I don't know how to run this any better than I know.
[00:12] SPEAKER_01: Okay.
[00:13] SPEAKER_01: Sally, first thing I want to ask is when you when you came to Sausalito?
[00:18] SPEAKER_00: Right after World War II.
[00:20] SPEAKER_01: Where had you been before?
[00:23] SPEAKER_00: Naval Air Station, San Diego, for the duration.
[00:29] SPEAKER_01: Where are you from?
[00:31] SPEAKER_00: I was born in Cleveland, Ohio.
[00:32] SPEAKER_01: Cleveland, Ohio.
[00:33] SPEAKER_00: April 30th, 1905.
```
Four things in those thirteen lines are worth more than a page of advice β and because the audio is right there, none of them has to be taken on faith.
"Now it's working" appears twice. It is not a bug in the transcription β it is on the tape. Someone is fiddling with a recorder that has just started working, and says it twice. A real transcript of a real interview begins with the equipment, not with the first question.
The machine gives you numbers, not names. `SPEAKER_00` and `SPEAKER_01` are what speaker detection produces. It can tell two voices apart and keep them apart for the whole recording; it cannot know which one is Sally. Assigning the names is a human step, and it takes about ten seconds β you find the first line where someone identifies themselves and rename both speakers everywhere.
"when you when you" is not a transcription error. The interviewer restarts the question, and the machine wrote both attempts down because that is what verbatim means. Play 00:13 and you will hear it. This is the moment where the style decision from the header becomes a real edit: verbatim keeps it, clean verbatim deletes one "when you", and a transcript that silently deleted it while calling itself verbatim would be misrepresenting the recording.
One line is doubtful, and you can see which. At 00:10, *"I don't know how to run this any better than I know"* is attributed to `SPEAKER_00`, the participant β but it reads like the interviewer still struggling with the recorder. This is exactly the kind of line to check against the audio before quoting. Attribution errors cluster in the first minute of a recording, where voices have not yet been heard enough to be told apart, and around interruptions.
Here is the same passage after the human pass: speakers named, the duplicate marked rather than deleted, the doubtful attribution flagged, and a proper header added.
```
INTERVIEW TRANSCRIPT
Project: Sausalito Historical Society oral history programme
Date: 5 January 1985
Location: Sausalito, California
Participant: Sally Stapp (b. 30 April 1905)
Interviewer: Sausalito Historical Society
Recording: Side A, 31 min 36 s (excerpt: 00:00β00:35)
Style: Verbatim
Source: California Revealed / Internet Archive, public domain
[00:02] INTERVIEWER: Now it's working. [repeated on tape] This is Sally
Stapp and it is January 5th, 1985.
[00:10] UNCLEAR: I don't know how to run this any better than I know.
[attribution uncertain β check audio]
[00:13] INTERVIEWER: Sally, first thing I want to ask is when you when
you came to Sausalito?
[00:18] STAPP: Right after World War II.
[00:20] INTERVIEWER: Where had you been before?
[00:23] STAPP: Naval Air Station, San Diego, for the duration.
[00:29] INTERVIEWER: Where are you from?
[00:31] STAPP: I was born in Cleveland, Ohio. April 30th, 1905.
```
Notice what changed and what did not. No words were altered. The duplicate was kept and annotated rather than silently removed, because in a verbatim transcript deleting it would be a small falsification. Sally's two consecutive lines were joined into one turn, since they are one continuous piece of speech that the machine split at a pause. And the doubtful line was marked rather than assigned to whoever seemed likely.
That is the whole job, and it is why an interview transcript is not simply the output of a transcription tool.
What an interview transcript is β and is not
A transcript records what was said, in the order it was said. It is not notes, not a summary, and not minutes.
The distinction matters most in research, journalism and any legal or HR context, because a transcript is the artefact you can be held to. If the transcript says something different from the recording, everything derived from it is compromised. This is why "clean it up as you go" is bad advice for a first pass: produce a faithful transcript first, then derive whatever you need from it.
The three transcript styles
Every style guide agrees on roughly the same three, and picking the wrong one wastes either your time or your data.
| Style | What it keeps | Use it for | Cost |
|---|---|---|---|
| Verbatim | Every word, filler, false start, repetition and pause; [laughs], [sighs], [pause] marked | Conversation analysis, discourse analysis, legal, anything where how something was said matters | Slowest to produce and hardest to read |
| Clean / intelligent verbatim | Every substantive word; filler and false starts removed, minor grammar fixed | Thematic analysis, most qualitative research, journalism, podcasts | Requires judgement about what counts as substantive |
| Edited | The core content; small talk and off-topic passages dropped entirely | Publication, internal reports, marketing case studies | No longer a complete record β never your only copy |
The practical rule: choose before you start, and write it in the header. A reader who does not know which style they are holding cannot tell whether the absence of hesitation means the participant was fluent or the transcriber was tidy.
How much filler is actually there?
The verbatim-versus-clean argument is usually conducted without numbers, so here are some. We measured filler across 934 English two-speaker recordings β the shape of a research interview β in our own production data:
| Token | Mean per 1,000 words | In a median 4,620-word interview |
|---|---|---|
| um | 2.16 | β 10 |
| uh | 0.92 | β 4 |
| you know | 5.20 | β 24 |
| like | 17.40 | β 80 |
Two honest caveats. "Like" is not a filler count β that 17.4 includes every ordinary use of the word, and we have not separated them. And these are means across recordings: filler is extremely uneven between speakers, so a single nervous participant can carry several times the average.
The useful conclusion is that cleaning filler out of a typical interview removes something like forty tokens from four and a half thousand words. It makes the transcript easier to read; it does not make it dramatically shorter. If you were choosing clean verbatim to save space, that is not a good reason. Choose it because your analysis does not depend on hesitation β and if it might, stay verbatim.
Interview transcript format: the three blocks
1. The header
Everything a future reader needs to interpret the text, before the text begins:
```
Project: [study or programme name]
Date: [date of the interview]
Location: [place, or "remote β Zoom"]
Participant: [name or pseudonym, plus any relevant detail]
Interviewer: [name or initials]
Recording: [file name, total length]
Style: [verbatim / clean / edited]
Anonymised: [yes / no]
Consent: [reference to the signed consent form]
```
The two fields people leave out and later regret are Style and Anonymised. Without the first, nobody can tell a fluent speaker from a tidy transcriber. Without the second, a reader six months later cannot tell whether "Maria" is a real person.
2. The body
The rules that actually matter, in order of how often they are broken:
1. One speaker per paragraph, always. Never merge two speakers into a block, even for a one-word answer. This is the single most common formatting failure and it makes a transcript unusable for quotation.
2. A consistent label followed by a colon. `INTERVIEWER:` and `STAPP:`, or `I:` and `P1:` β pick one scheme and never vary it. Analysis software splits on these labels.
3. Timestamps at a useful interval. Every speaker change is right for short interviews; every 30 to 60 seconds is enough for long ones. Timestamps exist so you can get back to the audio, and one every three seconds serves nobody.
4. Square brackets for everything that is not speech. `[laughs]`, `[pause]`, `[phone rings]`, `[crosstalk]`, `[inaudible 00:14:22]`. Include the timestamp inside an `[inaudible]` marker so it can be revisited.
5. Mark uncertainty rather than guessing. `[unclear: Kessler?]` is honest and useful. A confident wrong word is neither.
3. Line numbering, for research
If the transcript is going to be cited, number the lines. Qualitative work quotes as "P4, lines 212β218", and that reference is meaningless unless the numbering is stable. Add it once the transcript is final β renumbering after every correction is how numbers and quotations drift apart.
Interview transcript layout: what the data suggests
Layout advice is usually aesthetic. Some of it can be measured instead.
Across 1,248 two-speaker recordings of five minutes or longer in our production data, the median interview runs 29.9 minutes. In the 1,030 where we could measure talk time per speaker, the dominant speaker holds a median 72.4% of the talking, and at the ninetieth percentile 88.7%.
That number has two practical consequences for layout.
Your transcript will be lopsided, and that is correct. In a good interview the participant talks roughly three times as much as the interviewer. If your transcript is close to 50/50, that is worth noticing about the interview, not about the transcript. Our public-domain example above sits at 85.8% participant to 14.2% interviewer across the five-minute clip β near the top of the normal range, which is what you would expect from an oral history where the interviewer's job is to get out of the way.
Turns are short and numerous. The median interview in our data splits into 391 segments. Long unbroken paragraphs are almost always a formatting artefact rather than a feature of the speech, and they are the thing that makes a transcript painful to code. Break at natural pauses; do not let a single answer run for a page.
How to type a transcript of an interview
If you are transcribing by hand, the workflow that wastes least time:
1. Set the playback speed to about 75% in a player with keyboard control. Most people can type accurately at three-quarters speed but not at full speed, and constant rewinding costs more than slow playback.
2. Bind a pedal or a hotkey to rewind five seconds. Reaching for the mouse to scrub is the biggest single time sink in manual transcription.
3. First pass: get the words down. Do not format, do not fix names, do not look anything up. Type `[?]` and keep moving.
4. Second pass: fix the `[?]` markers, names, numbers and technical terms, with the audio.
5. Third pass: format β headers, labels, timestamps, line numbers.
Budget four to six hours per hour of audio for careful verbatim work. For the 29.9-minute median interview, that is two to three hours.
The faster route, and where the time actually goes
Automatic transcription does not remove the work; it inverts it. The machine produces the full text quickly and you spend your time reviewing, which is the part that was always the most valuable.
We measured turnaround on 962 completed two-speaker recordings: a median of 5.5 minutes from upload to finished transcript β about 6.5Γ faster than real time β with nine in ten finishing within 45.5 minutes. The long tail is mostly upload time on large video files rather than processing.
So for a typical 30-minute research interview: a few minutes to a transcript with speaker labels and timestamps, then your review pass. The review does not get faster, and it should not β it is where the accuracy comes from.
Practically, the sequence that works:
1. Upload the recording with speaker detection enabled.
2. Rename `Speaker 1` and `Speaker 2` to the interviewer and participant once, which updates the whole transcript.
3. Read it against the audio, correcting names, numbers, jargon and any overlap.
4. Apply your style: leave verbatim, or clean the filler.
5. Export as DOCX to edit, TXT for analysis software, or PDF to file.
Interview transcripts in qualitative research
A research interview transcript carries requirements that a journalistic one does not.
Anonymise consistently or not at all. Replace the participant's name, and also employers, colleagues, institutions and small places that would identify them. Use a consistent pseudonym or code throughout, keep the key in a separate access-controlled file, and record in the header that the transcript has been anonymised. Half-anonymised transcripts β pseudonym at the top, real employer in paragraph nine β are the common failure.
Keep the interviewer's questions in full. A leading question is only visible if it was recorded verbatim, and an answer often means little without the exact wording that prompted it. Since the interviewer accounts for only about 28% of talk time, keeping the questions costs little and removing them destroys the evidence of how each answer was produced.
Decide about filler before coding, not during. Switching from verbatim to clean halfway through a corpus makes the transcripts non-comparable, and any observation about hesitation becomes an artefact of when you changed your mind.
Check your software's import format. ATLAS.ti, NVivo and MAXQDA all take DOCX and plain text. Keeping a plain-text master avoids formatting surprises, and it is the format most likely to still open in ten years.
Common mistakes
Try it on your own interview
The transcript at the top of this page took about eighty seconds of machine time for a five-minute clip, with speaker detection on. You can run the same thing on your own recording β the first five minutes are free with no account, so you can judge the output on your own audio rather than on a demo we chose.
Upload an audio or video recording, let it detect the language, rename the speakers, and export as DOCX, TXT, PDF, SRT or VTT.
If your interview was recorded on video, MP4 to transcript covers what changes when the file is a video container. For recording technique and the habits that make a transcript easier before you press record, 10 pro tips from 10,000+ interviews is the companion to this page.
---
*Interview excerpt: "Oral Interview w. Sally Stapp", Sausalito Historical Society, 5 January 1985, via California Revealed and the Internet Archive (casauhs_000095). Public Domain Mark 1.0 β "Public domain. No restrictions on use." Transcribed 2 August 2026 from the first five minutes of Side A with speaker detection enabled; the raw excerpt above is unedited. Aggregate figures are from TranscribeNext production data, 9 October 2025 to 31 July 2026.*