Why a focus group is the hardest recording to transcribe
A focus group is the worst case for automatic speaker labeling. On clean one- or two-person audio, diarization error stays below 10%; on overlap-heavy, many-speaker recordings it runs 35-45% (Ryant et al., DIHARD III, 2021). More voices and more crosstalk mean more turns you'll fix by hand.
A typical group runs 6 to 10 participants (J-PAL, MIT), plus a moderator, all in one room. People finish each other's sentences, laugh over answers, and interrupt. That's exactly the signal that pushes a diarization model off track, because it has to guess who's speaking during every overlap.
The task underneath a good transcript is diarization: labeling who spoke each turn. It's reliable when voices are distinct and rarely collide. In a group of eight, the model still gives you a strong first draft. But the overlaps and the two soft-spoken participants at the far end of the table are where you'll spend your cleanup time.
How do you record so the voices stay separable?
Separation at the microphone is what saves you hours later. The model can only distinguish voices it can actually hear apart. Put a boundary or PZM mic in the center of the table for the group, and give the moderator a separate mic or channel, so the person steering the discussion is always cleanly captured.
Kill the room's noise before you record. Turn off HVAC, close windows onto traffic, and move away from the buzzing mini-fridge. A round table beats a long boardroom slab, because everyone sits roughly equidistant from the center mic instead of trailing off at the ends.
Open with a name round. Have each participant say their first name, or an assigned ID, in turn before the discussion starts. It anchors who 'Speaker 3' is, timestamps the moment for later, and gives you a clean voice sample of every person to match against.
How to transcribe a focus group into a clean draft
Typing a group discussion by hand is brutal: a single audio hour can take up to six hours of manual transcription (Haberl et al., 2023, citing Bell et al. 2018), and a 90-minute session is a full day gone. An AI first pass turns that into minutes of processing plus targeted cleanup.
Upload the recording to get a timestamped draft; turn on speaker detection, open to any signed-in account, to have each turn attributed to Speaker 1, Speaker 2 and so on. Read it against the audio and fix the predictable failure points first: the moderator's questions mislabeled as a participant, two people talking at once, and the quiet voices the mic barely caught.
Keep a roster as you go: a short list that maps each voice to its participant ID (for example, Speaker 3 is P3, who sat on the left) so every later edit is consistent. Correct the speaker labels in the transcript, and mark true overlap with a bracket, [crosstalk], rather than forcing a guess. A group draft will never be turn-perfect out of the box. Your job is to fix attribution wherever a quote or a coded segment depends on knowing exactly who said it. Crowded overlap needs a human pass, however good the tool. The verification steps in the qualitative research transcription guide show what to check and how.
What does a focus group transcript look like?
A focus group transcript lists each turn with a timestamp and a participant ID such as P2, keeps the moderator as M, marks overlapping speech as [crosstalk], and marks speech you cannot make out as [inaudible] with its time.
Here is a short excerpt, invented for illustration and not from a real session:
[00:12:04] M: What did you think when you first opened the box?
[00:12:09] P2: Honestly, I couldn't find the recipe cards, I had to dig under the –
[00:12:11] P4: [crosstalk] – the ice packs, yes.
[00:12:13] P2: Right, under the ice packs. [laughter]
[00:12:18] P1: Mine had the cards on top, so I never had that problem.
[00:12:22] P3: [inaudible 00:12:22] the box was bigger than I expected.
[00:12:27] M: P3, could you say that again?
Key: M is the moderator and P1 to P4 are participant IDs from the name round, never real names. A dash at the end of a line and the start of the next marks an interrupted sentence. [crosstalk] means several people spoke at once and the words cannot be separated. [laughter] records an event that matters to group dynamics. This is one convention, not a standard: pick one, write it down and use it throughout the file. The interview transcript template has a header and layout you can adapt, and the multiple speakers guide covers overlap in general.
| Marker | Use it when | Example |
|---|---|---|
| M and P1, P2, P3 | Every turn, so each line has a speaker | [00:12:04] M: What did you think? |
| [crosstalk] | Several people speak at once and the words cannot be separated | P4: [crosstalk] – the ice packs, yes. |
| [inaudible 00:12:22] | You cannot make out the words, with the time so you can return | P3: [inaudible 00:12:22] bigger than I expected. |
| A dash at the end and start of lines | A sentence is interrupted and finished by the same person | P2: I had to dig under the – |
| [laughter] | A non-verbal event matters to group dynamics | P2: Right, under the ice packs. [laughter] |
Is automatic speaker labelling the same as speaker identification?
No. Automatic speaker labelling separates voices and tags them Speaker 1, Speaker 2 and so on; speaker identification would work out who each person is by name. Pepys Transcribe does the first (for audio, at 1.5 credits per minute), not the second.
That is why the name round matters: it gives you a sample of each voice to match against "Speaker 3", and you then rename the labels to participant IDs. Speaker detection is open to any signed-in account, costs 1.5 credits per minute, and works on uploaded audio rather than on every link source. Pepys uses the voice analysis to produce the labelled transcript you asked for, not to identify people by name; speaker diarization explains how the labelling works.
What should you keep: verbatim, clean, or coding-ready?
Decide your transcription style before you edit, because in qualitative work the choice between naturalized and denaturalized transcription shapes the analysis (Oliver, Serovich & Mason, 2005), not just readability. Naturalized keeps every utterance in detail; denaturalized corrects grammar and strips interview noise.
For a focus group, the interaction is often the data: who agrees, who pushes back, where the laughter lands. If you're analyzing group dynamics, keep it naturalized, overlaps and all. If you only need the content of what was said, denaturalized reads cleaner. Pick one and apply it consistently across the file.
Then get it into your coding software. Export the cleaned transcript to DOCX for NVivo, ATLAS.ti, or Dedoose, keeping speaker labels and timestamps intact. For the deeper decisions, how to code group data and choose a convention, see the qualitative research transcription guide. For one-to-one interviews and the full path from consent to coding, see research interview transcription.
Consent, confidentiality, and anonymizing the transcript
You cannot promise confidentiality in a focus group. IRB guidance is blunt: because every participant hears each disclosure, the confidentiality of anything said in the session cannot be guaranteed (University of Connecticut IRB; Boise State IRB says the same). Tell participants that upfront.
Get consent to record, ideally captured in the audio. Recording-consent laws vary by state (RCFP): some need one party's consent, others require every participant to agree. With a roomful of people, all-party consent is the safe default, so collect a clear yes from everyone before you hit record.
Check whether your institution expects AI transcription to be disclosed. In a survey of 130 universities' public guidance, 25.6% of the 90 institutions with guidance said researchers must disclose AI transcription in the ethics protocol. The other 67 were silent or unclear, so that figure is a lower bound, and none were coded as not requiring it. Look up your own institution's rule rather than assuming either way. The study is a preprint of large research universities in the United States, United Kingdom and Canada, and it reads public web pages only.
Anonymize in the transcript, not just the report. Under the Common Rule's definition of identifiable private information (45 CFR 46.102), a name, employer, or a too-specific anecdote can re-identify someone. Replace them with role labels in a working copy and keep the master access-controlled. With Pepys, the source recording is auto-deleted about 30 days after upload, while the transcript and your exports are kept. The sensitive original file doesn't linger on a server.



