Focus Group Transcription: How to Transcribe a Focus Group (and What to Expect from AI)

Why are focus groups the hardest audio to transcribe?

Focus groups are hard to transcribe because several people talk to each other rather than to one interviewer. Voices overlap, short reactions fly across the table, participants sit at different distances from the microphone, and the transcriber has to work out not only what was said but who said it.

The whole point of the method is interaction, so the richest moments are often the messiest on the recording: two participants disagreeing at once, a burst of laughter, someone finishing another person's sentence. Julia Bailey's methods paper on transcribing notes that transcription takes at least 3 hours per hour of talk, and up to 10 hours per hour when a fine level of detail is needed. Group recordings, with their overlapping voices, usually sit towards the slow end.

The difficulties fall into five groups:

• Overlapping speech: people talk over each other when the discussion gets lively
• Similar voices: several adults of the same age and accent are hard to tell apart on audio alone
• Uneven distance to the microphone: the nearest person is loud, the one at the far end faint
• Back-channel talk: "yeah", "exactly" are short and easy to misattribute, yet they can signal agreement you want to analyze
• Moderator versus participants: prompts must be clearly separated from the group's answers

The quality of your transcript is therefore decided long before anyone types a word, when you plan the recording.

How should you record a focus group to make transcription easier?

Record a focus group with a good central microphone, a backup recorder, a seating chart and a round of introductions. These habits cost minutes on the day and save hours of guesswork afterwards.

Microphone placement: one mic or several?

One omnidirectional conference microphone in the middle of the table works for a small group seated close together. For a long table or soft-spoken participants, add a second recorder at the other end, or individual lapel microphones if your setup allows: separate channels make attribution far easier but add a synchronization step. Either way, run a short sound check with the actual group.

For online focus groups on Zoom, Teams or Meet, use the platform's recording or capture your computer's system audio, and ask participants to wear headsets and mute when not speaking. A 2024 methods paper by Eftekhari on speech recognition in qualitative research recommends good quality audio, headsets to minimize echo, less overlapping speech and a quiet environment to limit transcription errors.

Video, seating chart and introductions

• Video as a speaker aid: a camera in the corner helps you see who spoke during a heated exchange. If you film, make sure your consent process covers video and keep it only as long as you need it.
• A seating chart: note each participant's code (P1, P2...) and seat. The note-taker can also jot down the first words of key contributions with a time stamp.
• Introductions on tape: ask each participant to say their code or first name and one short sentence. You get a clean voice sample for every person, invaluable when you match voices to labels.
• Ground rules: a friendly "one person at a time, please" will not eliminate crosstalk, but the moderator can come back to it when the discussion overheats.

Full verbatim or clean verbatim for a focus group?

Choose clean (intelligent) verbatim if you plan a thematic or content analysis, and full verbatim if the interaction itself is your data. For most applied, market and UX projects clean verbatim is enough; studies of how opinions form in the group, or of how people talk, need full verbatim with overlap marked.

Even with clean verbatim, do not strip the short reactions that carry meaning in a group. "Yeah, same" from three participants is evidence of shared experience, and the Eftekhari paper lists square brackets as a common convention for marking overlapping talk or emotion. For the two styles in depth, see our guide to verbatim vs clean verbatim transcription; for transcription across qualitative methods, see our guide to transcription in qualitative research.

How do you label and anonymize speakers?

Label every turn with a stable code, such as M for the moderator and P1 to P8 for participants, and keep the key that links codes to real people in a separate, secured file. Codes are easier to anonymize than names and stay consistent across groups.

• Prefix codes by group when you run several sessions (G1-P3, G2-P3) so participants from different groups never share an identifier.
• Use "Unidentified" rather than guessing. If you cannot tell who spoke, write [P?] or [Several]: a wrong attribution silently distorts any analysis by participant.
• Watch what your labels imply. The Eftekhari paper recommends checking whether a naming convention maintains pseudo-anonymity or ascribes a role, gender, identity or power relation.
• Replace identifiers consistently. The UK Data Service guidance on anonymizing text data advises replacing names, employers and specific locations with pseudonyms or descriptive tags such as [colleague 1], applying replacements consistently across transcripts, and keeping an anonymization log of what you changed.

In a group, participants also name each other ("As Sarah said..."): replace those names with the matching code too.

How to transcribe a focus group step by step

The fastest reliable workflow is an AI first pass followed by a careful human review: the machine types, you listen and judge.

How long does focus group transcription take, and what does it cost?

By hand, expect several hours of work per hour of recording, and more for a busy group with overlapping voices. With an AI first pass, the typing disappears but the review does not.

On cost, doing it yourself costs only your time, which the figures above show is considerable. If you request quotes from a human transcription service, check whether the price changes with the number of speakers, the verbatim level, speaker identification and turnaround, since focus groups tend to need all of them. An AI tool plus your own review saves typing time, not thinking time.

What can AI do with crosstalk, and what can't it?

AI handles the clear, one-at-a-time stretches of a focus group quickly and well, but overlapping speech remains its weak spot: expect missing words, merged turns and misattributed speakers where people talk at once.

This is not a limitation of one product. The review of speaker diarization research by Park and colleagues describes speech recognition on overlapping speech as one of the main challenges in meeting transcription, reports that studies of meeting recordings observed an average of 12% to 15% of speaker overlap, and notes that many traditional diarization systems focused only on non-overlapping regions. A lively focus group is rarely an easier case.

In practice, an AI transcription tool will:

• Do well: produce a fast, time-stamped draft of clear speech, separate distinct voices into labeled turns, and save you the hours of typing
• Struggle: keep both sides of a simultaneous exchange, catch quiet back-channel responses, tell apart two participants with very similar voices, and spell specialist jargon or brand names
• Not do at all: know who each person is. Voice separation tells you that Speaker 3 is a different voice from Speaker 5; only you, with your seating chart and introductions, can say that Speaker 3 is P4

Consent, confidentiality and GDPR for focus group recordings

Recording a focus group needs every participant's informed agreement, a warning that confidentiality within the group cannot be guaranteed and, in the EU and UK, a lawful basis and safeguards for processing the recordings.

The University of Connecticut IRB guidance on focus groups sums up the specifics: disclose the use of recording devices in the consent form and consent discussion, ask each member again at the session whether they agree to be recorded, tell participants that what is said should not be shared outside the group and that the confidentiality of what they say cannot be guaranteed, and ask them not to use names when the session is recorded.

Data protection is a separate question from research ethics. The UK regulator's guidance on the research provisions of the UK GDPR points out that consent to take part in a study is distinct from consent as a lawful basis for processing, and that in most research cases consent is not the most appropriate lawful basis, partly because it can be withdrawn at any time. In a group recording, think in advance about what a withdrawal would mean in practice, and agree the lawful basis with your data protection officer or ethics committee.

Finally, check where your transcription tool processes and stores the audio: a recording of eight voices is personal data about eight people. A processor hosted in Europe that deletes the audio after processing keeps the data trail short for European studies.

How the transcript feeds thematic analysis

A focus group transcript is analyzed at two levels: what was said (themes across the group) and how it was said (agreement, disagreement, how views shifted in discussion). Accurate speaker labels are what make both possible.

For thematic analysis, the approach developed by Virginia Braun and Victoria Clarke is the usual reference point: a family of methods for exploring and interpreting patterned meaning across a dataset. With focus groups:

• Code at the group level first, then check whether a theme is carried by the whole group or by one or two dominant voices.
• Keep speaker codes as attributes in your analysis software so you can filter contributions by participant, group or demographic.
• Code the moderator separately, or exclude their turns, so prompts are not mistaken for participant views.
• Mark consensus and dissent. Short reactions you preserved in the transcript ("same", "no, not for me") are what let you show agreement rather than assert it.

Common focus group transcription mistakes

1. Relying on one recorder. A single device with a flat battery or a muffled corner loses a session you cannot rerun.
2. Skipping introductions. Without a clean voice sample per participant, matching voices to codes becomes guesswork.
3. Publishing the AI draft. An unreviewed draft is fine for your own notes, never for coding or quotes.
4. Guessing speakers. Mark uncertain turns as uncertain instead of forcing an attribution.
5. Deleting group reactions in the name of clean verbatim, and with them the evidence of consensus.
6. Mixing conventions across groups. Write your transcription rules down before the first session and stick to them.
7. Forgetting names spoken aloud during anonymization, including names of colleagues, employers and places.

Where AudiosTranscribe fits

AudiosTranscribe provides the AI first pass described above: a time-stamped transcript with voices separated into distinct speakers, ready for your review.

For one-to-one interviews, see our interview transcription page; the pricing page details each plan.

Sources and further reading

• Bailey, J. (2008). First steps in qualitative data analysis: transcribing. Family Practice, 25(2), 127 to 131.
• Eftekhari, H. (2024). Transcribing in the digital age: qualitative research practice utilizing intelligent speech recognition technology. European Journal of Cardiovascular Nursing, 23(5), 553 to 560.
• Park, T. J., Kanda, N., Dimitriadis, D., Han, K. J., Watanabe, S. and Narayanan, S. A Review of Speaker Diarization: Recent Advances with Deep Learning Computer Speech & Language, 72, 101317 (2022); preprint arXiv:2101.09624.
• UK Data Service. Anonymisation for text data.
• University of Connecticut, Office of the Vice President for Research. IRB researcher guide: focus groups.
• Information Commissioner's Office. The research provisions: principles and grounds for processing.
• Braun, V. and Clarke, V. Thematic analysis (the authors' resource site).

Your institution's ethics committee or IRB may set stricter rules, so check its requirements before your first session.