Interview transcript cleaner

Automatic transcription has made the recording-to-text step nearly free, and moved the work downstream. Every meeting tool exports its transcript in a different shape, built for re-watching a meeting rather than for analysing an interview. The cleaner above reshapes those exports into the form qualitative analysis expects; this page explains the choices it makes and the ones it leaves to you.

What each tool gives you

• Zoom cloud recordings produce an audio transcript as a WebVTT (.vtt) file: a caption cue every few seconds, each with a start and end time and the speaker's display name before a colon. One answer can span a dozen cues.
• Microsoft Teams offers the transcript as .vtt or .docx. Its VTT marks speakers with voice tags (<v Name>) and gives each cue a long identifier; its Word export puts the name and a time on a line above each paragraph, and adds "started transcription" lines that are not speech.
• Otter exports TXT with a header line per paragraph—"Speaker 1  0:07"—and splits long turns into several paragraphs under repeated headers.
• Rev and similar services write "Name (00:01):" before each turn.

The cleaner detects which of these it has been given and says so, then rebuilds the conversation as one paragraph per turn. If a tool exported captions without speaker names—Zoom does this when speaker identification is off—it keeps the text but cannot split the speakers, and tells you.

Verbatim, clean verbatim, and what to report

True verbatim keeps every filled pause, stutter and false start. Clean verbatim removes those disfluencies while keeping every content word and the speaker's own phrasing—the convention most thematic and content analyses use, and the default here. The readable preset also removes hedges set off by commas (", you know,") and sound tags such as [laughs], which some teams prefer for sharing quotes and others consider a loss of meaning. Each rule is separate, and the "what changed" view colours every removal by the rule that made it, so the decision is visible rather than buried.

Which convention is right depends on the analysis. Conversation analysis, discourse analysis and studies of hesitation or speech in psychosis treat disfluency as data and need true verbatim—often with more detail than an automatic transcript contains. Clinical rating from interviews usually works from clean verbatim. Whichever you use, say so in the methods, along with who checked the automatic transcript against the audio.

Speaker labels

The cleaner takes the speaker who asks the most questions per word as the interviewer and labels them I, with participants as P (or P1, P2 in group interviews). Labels are yours to change, and they replace the display names in the output—which matters, because a Zoom display name is often the participant's real name. Consistent labels at the start of each paragraph also let NVivo and MAXQDA auto-code the transcript by speaker, and are what Tagaroo's importer reads.

Before the transcript leaves your machine

Interview transcripts are among the most identifying data a study holds. The de-identify option replaces names, contact details and dates with consistent pseudonyms, using the same detector as our transcript de-identifier, and applies it across the whole transcript so that one person keeps one pseudonym. It only applies high- and medium-confidence detections automatically; for a file you are about to share, run it through the de-identifier's review screen, which lists every detection including the uncertain ones.

How accurate was the automatic transcript?

Cleaning does not fix recognition errors. If you have corrected a sample of a transcript by hand, the word error rate calculator compares the two and shows which kinds of error the tool makes, and the transcription time calculator estimates how long checking the rest will take. Our guide to preparing audio and speech for annotation covers the upstream choices—diarisation, overlapping speech and transcription conventions—in more depth.

What this tool does not do

It does not transcribe audio; it works on the text your tool already produced. It cannot recover speaker names that were never recorded, and it cannot fix a turn the recogniser attributed to the wrong person—check speaker changes against the recording where they matter. The filler rules are English-language and deliberately conservative: they remove "um", "uh" and "erm", immediate repetitions and cut-off words, and leave "like" and "you know" alone unless they are set off by commas, because in the middle of a sentence they are often meaningful.

Frequently asked questions