A one-hour interview has two voices taking turns. A one-hour focus group has a moderator, six or eight participants, and a few moments where three people talk at once because someone said something worth reacting to. Those moments are usually the ones you wanted. It's also why focus group transcription fails in a different way from ordinary interview transcription, and why most of the fix happens before anyone presses record.

We sit on both sides of this at Spirelight. We run human transcription for speech projects, and we built a self-serve transcription tool. My short answer: an automatic draft is a fine place to start, but a group recording won't give you trustworthy speaker labels on its own. Budget for a human pass on who said what, and spend your real effort on the recording.

Why focus group transcription is harder than an interview

Two separate jobs happen in any transcript. One is getting the words right. The other is deciding who said them. A group makes both harder, and the second one suffers most.

Most transcription tools assign each word to a single speaker. When two people overlap, the tool has to pick, and the quieter voice can vanish from the text altogether. Short interjections like "exactly" or "no, but" get glued to the wrong person. Two people with similar voices can merge into one speaker halfway through. We've written up where speaker labels break in more detail for engineering teams, and most of that list shows up in a focus group.

Published accuracy numbers flatter you here. A 2023 study in JMIR Mental Health compared services on 65 one-to-one research interviews. Median word error rate was 7.6% for a human service, 8.9% for the best automatic one, and 14.8% and 19.2% for the next two. Read the setup, though. The automatic services were given only the interviewee's audio track, and the authors note there was relatively little background noise. A focus group is close to the opposite of that, so treat those numbers as a best case you won't get.

Closer to our situation, researchers at CISPA in Germany tested eleven providers on interviews and group discussions in German and English, some with background noise added. Manual services came out ahead, and the team reported that AI services often produced errors that distorted the meaning. Their example is a good one: "hashes" became "ashes". That test dates from December 2022 and the tools have moved on since, but the lesson holds. One wrong word in the wrong sentence costs more than an error rate suggests.

Record it so the transcript has a chance

No transcriber, human or machine, can recover a voice that never reached the microphone. If you can only change one thing, change this.

For an in-person group, put two recorders on the table instead of one, at opposite ends, and keep cups and phones away from them. Ask each person to say their participant number once at the start, one at a time, so there's a clean sample of each voice and no name in the recording. For an online group, record separate audio tracks if your platform offers it. Zoom does. One speaker per track removes most of the labelling problem with a single setting.

Then give someone a job the recording can't do. Robert Brunger's practical tips for focus groups are from 2010 and still right: have a note-taker draw a seating diagram, give each person an identifier, and start a stopwatch with the recorder so the notes can be matched to the audio later. One addition: during heated stretches, have the note-taker jot down the first few words of each turn. "P4: that's not what" at 23:10 will settle an argument about a speaker label in seconds.

Decide what the transcript is for

Before you pick a provider, decide how you'll analyse the group, because that sets how much detail you need. A thematic analysis can live with cleaned-up text, as long as the speakers are right. If you're studying how people build on or interrupt each other, you need overlaps, false starts and pauses marked, and that's the detail automatic tools tend to smooth away.

Whatever you choose, write it down. The UK Data Service's guidance on documenting data is blunt about it: "Transcription conventions must be documented." It also asks for speaker tags to be applied consistently, which matters more in a group than anywhere else. Here's a fictional excerpt in a format that holds up:

[00:14:32] MOD: What made you stop using the app?
[00:14:36] P3: Honestly, the price.
[00:14:37] P5: Yeah, same.
[00:14:38] [crosstalk, 3 sec]
[00:14:41] P?: ...but you get it free through work, right?
[00:14:43] P3: Not anymore.

Three details do the work there. A time code on each turn lets anyone jump back to the audio. A fixed set of labels (MOD, then P1 to P8) keeps real names out of the file from the start. And P? admits doubt instead of guessing, so a reader can see that line still needs checking.

AI draft, human transcriber, or both?

Cost is why people ask. In the JMIR study, the human service cost US $1.50 a minute and the best automatic one US $0.024. Prices vary, but the gap is real, so the useful question is which parts of the job need a person.

If you have a handful of groups, decent audio and a thematic analysis, start with an automatic draft and fix it yourself with the seating notes beside you. That's the job we built Spirelight Transcription for: you get a draft with speakers and time codes, correct it while you listen, and code it in the same project. Test any tool on a harmless clip first. Standard Spirelight Transcription accounts don't include a data processing agreement, so if your participants can be identified, ask your institution before you upload anything, to our tool or anyone else's. Loughborough University's data privacy team warns that many services use uploaded audio to train their AI models, so check that for every provider you consider.

If the groups are sensitive, run in a language or dialect the tools struggle with, or sit at the centre of a thesis or a paid study, use a human transcriber and hand over your seating notes. Mixed setups work too: AI for the words, then a person for the labels.

How to check a focus group transcript

You don't need to re-listen to everything to learn whether you can trust it. Pick three short passages and check them against the audio with your notes open:

  1. A disagreement, where people talk over each other.
  2. A fast run of short replies.
  3. A moment from your quietest participant.

Those are the places labels fail first. If they're right, the easier stretches probably are too. If they're wrong, assume the problem runs through the whole file. And before you quote anyone in your findings, listen to that line yourself, whoever made the transcript.

If your groups run across languages, or the speaker labels have to hold up under someone else's scrutiny, that's the point to bring in people. Our human transcription service covers 13 languages through native transcribers, and you can start with a free sample.