How to Transcribe Research Interviews Properly: A Step-by-Step Workflow

Qualitative research lives or dies on the quality of its transcripts. You might be coding interviews in NVivo or ATLAS.ti, quoting participants in a thesis, or trying to make sense of twelve hours of fieldwork. Either way, you need text that is accurate, labelled by speaker and timestamped, so you can search it and cite it. You also need to be able to defend how it was made.

"Transcribing properly" means making a few decisions before you start typing or uploading, and doing a few checks after. This guide walks through them in order. It uses an AI first draft and a human verification pass, which is how most researchers work now.

Before you transcribe: three decisions

1. How verbatim does it need to be?

This depends on your analysis method, so decide it first.

Level	What it keeps	Fits
Clean verbatim	Every meaningful word. Drops "um", false starts and stutters.	Thematic analysis, content analysis, most applied and policy research
Verbatim	Everything said, including fillers, repetitions and false starts, with [pause] and [laughs] marks	Studies where hesitation and emphasis matter, such as sensitive topics or narrative analysis
Detailed notation	Overlaps, pause lengths, intonation, using a system such as Jefferson notation	Conversation analysis and discourse analysis

AI transcription gives you something close to clean verbatim. Whisper-based engines, ours included, leave many fillers and false starts out. That's ideal for thematic work. If you need full verbatim, plan to add the disfluencies by hand while listening. Detailed notation is always a manual job, done on top of a transcript.

Write the level you chose into your methods section. Reviewers ask.

2. How will you label speakers?

Use roles or codes, not names: I: for the interviewer, and P01:, P02: for participants. The code is the anonymisation, so keep the key that links P01 to a real person somewhere separate from the transcripts. In focus groups, ask each person to say their code once at the start of the recording ("I'm P03"). Later, that one sentence lets you match each voice to its label in seconds.

3. What notation will you use for problems?

Agree on a small set of marks and use them the same way in every transcript, so the transcripts can be searched and compared:

Situation	Mark
Can't make out a word or phrase	[inaudible 00:12:31]
Your best guess at an unclear word	[sourdough?]
Two people talking at once	[crosstalk]
Laughter, sighs, long pauses	[laughs] [sighs] [pause]
Someone cuts in	End the line with --
Removed for anonymity	[employer] [city] [colleague's name]

Always put a timestamp on [inaudible]. Then anyone, including you in three months, can go back to that exact moment and try again.

Recording for a good transcript

A close microphone and a quiet room do more for accuracy than any software setting.

• One mic per person if you can. Separating the voices in the audio is what lets the software separate them in the transcript.
• Record a ten-second test and listen back before the interview starts.
• Don't talk over each other. Overlapping speech is where both the words and the speaker labels go wrong. Wait a beat before you ask the follow-up.
• Keep a running list of names, places, acronyms and technical terms the participant uses. You'll need it for the checking pass.
• Save as wav, mp3 or m4a. Any of these works.

The workflow, step by step

1. Upload and turn on speaker detection. In ScribeToAny, open the extra settings, tick Detect speakers and set how many people are talking. For a one-to-one interview that's two. Leave the language on auto-detect (about 98 languages are supported), or choose it yourself if you know it.
2. Read the draft once, without stopping. Get a sense of where it's weak: crosstalk, a quiet participant, a stretch of jargon.
3. Do the checking pass with the audio. Go through it again while listening. Clicking a timestamp replays that moment, and clicking a line lets you fix it. Correct names and terms from your list, and add your notation marks as you go.
4. Check the small words. AI mistakes are rarely big. In a scripted two-person test we ran, the transcript got all 14 speaker turns right, but it turned "the one thing you'd tell" into "you tell". That one-word difference changes what the speaker means. If you'll quote a passage, always check the exact wording against the audio. (The full test is in our real transcript examples.)
5. Rename and anonymise. The transcript labels voices SPEAKER_00, SPEAKER_01 and so on. Once exported, use find-and-replace to change them to I: and P01:. Then replace any identifying details with the bracketed marks above.
6. Export for your analysis tool. Use DOCX or TXT for reading, quoting and importing into NVivo or ATLAS.ti as documents. Use TSV or CSV for a spreadsheet with one row per segment and columns for start time, end time, speaker and text. Use JSON if you're scripting the analysis.
7. Name files consistently. For example, P01_2026-09-14_interview.docx. Store them with your other study data, under your ethics approval's data-handling rules.

Why the human pass isn't optional: automatic transcription is a first draft, not a record. For anything you'll quote or code, check the passage against the original audio. Good rigour and good research ethics both require it, and it's the step your methods section should describe.

Focus groups and group interviews

More voices make everything harder. Diarization is the step that works out who spoke when. It struggles with similar voices, short replies ("yeah", "right") and people talking over each other. (See what speaker diarization is and where it fails.) Three things help:

• Set the speaker count to the real number of people instead of leaving it on auto.
• Use the start-of-recording introductions to match each voice to its code.
• When a two-word reply can't be attributed, label it [unclear speaker]. Don't guess.

Non-English interviews

If you interview in a language other than English, ScribeToAny transcribes in the source language on every plan, including the free one. You can also translate a finished transcript into any of 134+ languages for free. That's useful when you need an English working copy alongside the original-language transcript. Quote from the original, and state in your methods section that the translation was machine-generated and how you checked it.

Privacy and consent

Interview data is sensitive. Only transcribe recordings you have consent to process. Check that your consent form covers the use of an automatic transcription service. Follow your institution's data-handling and ethics-board rules, and delete source files you no longer need. Uploaded files are used to produce your transcript, not to train models.

How to cite ScribeToAny

If a transcription tool shaped your data, name it in your methods section. A general form:

Interviews were transcribed using ScribeToAny (https://scribetoany.com), an automatic speech-to-text service, in clean verbatim, and corrected by the author against the original audio.

APA-style reference:

ScribeToAny. (2026). ScribeToAny [Automatic transcription software]. https://scribetoany.com

Adjust it to your discipline's style guide. What matters is a statement that someone else could repeat: what produced the transcript, at what level of detail, and that a person checked it.

The free plan for a small study

The free plan covers 2 transcriptions a day, 100 minutes a month and 30 minutes per file, with no credit card. That's enough for a pilot or a handful of short interviews. For a full study with long sessions, the paid plans raise the per-file and monthly limits. If you're budgeting for a grant, our transcription pricing comparison works out the cost per hour of audio here and at Otter, Rev, Sonix and HappyScribe.

---

Decide the verbatim level, the speaker codes and the notation first, and treat the AI transcript as a draft you check against the audio. Then your transcripts will be ready to code, not ready to redo. Start with the interview transcription tool.