How to transcribe a focus group, and where automatic transcription stops

Transcribe a focus group and you are handing an automatic tool the hardest job there is. Six or eight voices, constant overlap, and usually one microphone in the middle of a table. You will get a transcript back, and it will be a rough draft rather than a finished artefact: no speaker labels, and mush wherever people talked over each other. What decides how usable it is happens before you record and after the tool runs.

Overlap is the core problem, not accents and not noise

A speech model reads the audio in fixed slices. The Whisper paper puts it plainly in section 2.1: “We break audio files into 30-second segments paired with the subset of the transcript that occurs within that time segment.” One window in, one stream of text out. There is no branch in that output for a second voice, so when two people speak across the same seconds the model writes one sentence where two were spoken. It picks one, blends them, or skips the moment.

That makes overlap a different kind of problem from the room and the accents. Those degrade a signal that still contains one person talking, and a closer microphone and a quieter room fight both. Overlap is not a degradation. Both voices are in the file; the output has nowhere to put the second one.

There is no error rate to quote for it. The Whisper paper does not address overlapping speech anywhere, and we have not measured focus groups either, so there is no number here. What the paper does describe, in section 6, is what goes wrong in long recordings: “problems such as getting stuck in repeat loops, not transcribing the first or last few words of an audio segment, or complete hallucination”. A ninety-minute session with eight people is exactly that case. Our post on recording audio that transcribes well helps with the ordinary problems, but not this one.

What you get back when you transcribe a focus group

Plain text and timings. That is the whole output. No names, no “Participant 3”, and nothing marking a change of voice.

The missing labels are deliberate. Section 2.4 says the models had “a tendency to transcribe plausible but almost always incorrect guesses for the names of speakers”, and the authors fine-tuned that out “on the subset of transcripts that do not include speaker annotations”. There is no setting to turn on, and our post on what speech recognition does and does not do with two speakers covers why attribution is a separate task. So you cannot count talk time, and you cannot attribute a quote without going back to the audio. The passages you most want, where the group argued, are the ones where the overlap is worst.

The rest of the limits matter here more than anywhere else. FreeTranscribe runs OpenAI’s open-source Whisper model in the base English size on your graphics card through WebGPU, so it needs desktop Chrome or Edge with a working WebGPU adapter and it is English only for now. The model is about 200 MB, downloaded once and then cached. Names, technical terms, heavy accents and noisy rooms are its weak spots. Nothing is uploaded and it exports TXT, SRT and VTT. In our test it ran at about 1.5x real time on a desktop with a graphics card, and roughly the length of the recording on a thin laptop.

More microphones is the single biggest change you can make

Everything else here is damage control. This is the one decision that changes the problem.

Record each participant to their own track and overlap stops being unrecoverable. Two people talking at once become two files, each with one dominant voice, and you transcribe them one at a time. The other voices still bleed in, but the sentence buried in the room mix sits in the open on its owner’s track. Attribution comes free, because a track belongs to a person before you transcribe a word of it.

Recording setup	What it buys you	What it costs
One recorder in the middle of the table	Nothing beyond a record of the session. Every attribution is done by ear later	Cheapest and worst. Overlap is unrecoverable
Two recorders, opposite ends of the table	A second angle to listen to when the centre mix turns to mush	One more device, and two files to line up
A clip mic per participant into a multitrack recorder	One track per person, so overlap separates and attribution comes free	The kit, the setup time, and a longer transcription job per session
A phone per participant, recording locally	Close to per-person tracks for the price of nothing	Syncing the files, battery and storage, and asking each person’s permission
A boundary mic per three or four people	Smaller groups per track, so collisions inside a track drop	A mixer or a recorder with several inputs
Video on the group as well as audio	Faces tell you who spoke when the audio will not	Another consent conversation and another file to store

Six tracks means six runs through the tool, and six transcripts you interleave by start time. Start every device, then clap once, so you have a reference point when you line them up.

The moderator decides how transcribable the recording is

Nobody will do any of this unless the moderator asks, and asking costs about forty seconds at the top of the session.

Open with a round of names in each person’s own voice. It is the only thing that keeps a voice identifiable weeks later.
Say the session is being recorded and ask for one voice at a time. People comply when the reason is concrete.
Restate the question before each new answer. That gives you a searchable anchor and a marker for where a turn began.
Name the next speaker. “Sam, then Dan.” Those three words are a speaker label written into the audio itself.
Nod instead of agreeing out loud. A moderator’s “mm-hm” over the middle of an answer costs you that window.
When two people collide, ask one to repeat the point. One clean repetition beats any editing afterwards.
Leave a beat at each turn change. A pause gives the model a clean boundary, so cue breaks fall between speakers rather than through them.

A disciplined group of eight transcribes better than four people arguing. If yours is the argumentative kind, add microphones rather than removing people.

Adding attributions by hand is a listening pass, not an editing pass

You are playing the recording and writing down who is talking, and nothing removes the playback.

Export SRT or VTT, not TXT. You need the timings to navigate, and our post on subtitle and transcript formats explains the difference.
Fix names and vocabulary with find and replace first. The model gets these wrong consistently rather than randomly, so one replacement fixes every instance. Transcript mistakes to fix before you publish covers the rest.
Play it through and type an initial at the start of each cue. One character per turn, identical every time so find and replace still works later.
Flag the collisions, do not fix them in place. Drop a marker like [CROSSTALK 00:14:22] and keep moving. Untangling each one as you reach it is how a two-hour job becomes a two-day one.
Come back to the flagged spots last. With per-person tracks, open the track belonging to whoever you think was speaking. Otherwise mark it inaudible, because guessing at a quote is worse than losing it.
Only then start coding. An unattributed transcript is not ready for analysis.

Budget for playing the whole recording at least once. That floor is set by the length of the session, not the length of the text, and we have not measured it across real sessions.

Paying a human transcriptionist is often the right answer here

This is the one case where we would tell you to get your wallet out. Our tool gives you words and timings from a hard recording, not attribution, and no setting changes that.

Pay someone when the transcript is a deliverable rather than a working note, or when the analysis depends on who said what, which for most focus group work it does. Ask how they price per audio hour, whether speaker attribution is included, and how they mark crosstalk. Our reference on transcription cost per minute covers the money side.

A human hears the same overlapping audio you do, and can rewind and use context our tool has no access to, but nobody can un-mix two voices, so a better recording pays either way. And sending the file puts one more copy of your participants’ voices outside your control, which our post on personal data in meeting transcripts is a starting point for.

For sessions you work through yourself, or for a rough first pass before you decide whether to pay for a clean one, drop the recording on the home page and export SRT.

Frequently asked questions

Will anything label the speakers in a focus group recording? Separating audio by voice is a different task from recognising words, handled by different software, and FreeTranscribe does not do it. We have not tested diarization tools on a room of eight, so we will not tell you how any of them perform. What we can say is that one mixed recording holds the same overlapping audio whatever reads it.

Would a bigger model handle the overlap? No. Section 6 of the paper notes that larger models made “steady and reliable progress on reducing perception-related errors such as confusing similar-sounding words”, and overlap is not that kind of error. Our post on speech model sizes covers what size does and does not change.

Do I need to transcribe the whole session? If you are coding systematically, yes. If you want a handful of illustrative quotes, a marked-up listen plus accurately typed quotes is a legitimate output, and the same logic applies as in our guide to transcribing an interview.

Sources, checked 15 September 2026

https://arxiv.org/html/2212.04356 - Radford et al., “Robust Speech Recognition via Large-Scale Weak Supervision”: section 2.1 (30-second segments), 2.4 (incorrect guesses at speaker names and the fine-tuning that removed them), 6 (larger models reduce perception-related errors; long-form failures including repeat loops and hallucination), and the absence of any treatment of overlapping speech.
The only source fetched. Recording setups, moderation and the attribution pass are our own practice, and FreeTranscribe’s formats, limits and speed come from our own testing.