Focus Group Transcription

Most of the accuracy in a focus group transcript is decided in the room, before anything is uploaded.

1. Put one good microphone in the middle of the table, not a laptop at one end. The people furthest from the recorder are the ones who come back as fragments.
2. Ask everyone to say their first name, or their participant number, once at the start. That gives you a clean sample of each voice, so you can match "Speaker 3" to P3 in a minute rather than by guesswork.
3. Ask the moderator to slow down crosstalk when it starts. Two people talking at once is the single biggest source of lost text.
4. Record the whole session as one file. A 90-minute group uploaded in one piece keeps one consistent set of speaker labels, where three separate files would each start again from Speaker 1.

Video recordings from Zoom, Teams or a camera work too: upload the video file and the audio is taken from it.

Each turn in the transcript starts with a timestamp in square brackets and a speaker label, for example "Speaker 2:". That structure is what makes the next steps quick.

• Name the speakers. Use find and replace in the Word or text file to turn "Speaker 2" into a pseudonym or participant code. Do it once, before anyone starts coding.
• Anonymise. Participants say names, employers and places out loud. Search the transcript for them before it leaves your team; the timestamp beside each turn takes you back to the audio when a word is unclear.
• Import. The plain-text file goes straight into NVivo, ATLAS.ti, Dedoose, MAXQDA or a spreadsheet. The Word file keeps speaker names in bold, which is easier for a second coder to read.
• Check the quotes you publish. Read any quote you will use in a report against the audio. Automatic transcripts are close, not certified, and a dropped "not" changes a finding.

The first 10 minutes each day are free, enough to check that your recording works before you pay for the whole session. After that it is $0.05 a minute (1 credit), billed on the length of the recording. A 90-minute group is 90 credits. Credits come in one-time packs and never expire, which suits research that runs in bursts.

Files up to 500 MB are accepted. The recording is deleted from storage as soon as the job ends, and the transcript is kept for 24 hours, or 30 days if you choose, with a button to delete it sooner. Nothing is published. If your ethics approval names the services that may process recordings, the security page lists them.

Focus group transcripts are read differently from interviews: you are looking for patterns across speakers rather than quotes from one, and the unit of analysis is usually the turn.

That makes the paragraph-per-turn structure here directly usable. Export plain text and import it into NVivo, Atlas.ti, Dedoose or a spreadsheet; each turn keeps its timestamp in square brackets, which survives the import and lets you return to the audio for anything ambiguous.

Be realistic about speaker attribution with six or eight participants around a table. Diarisation will merge similar voices and split people who move. For research where who-said-what is load-bearing, take a seating note during the session and reconcile the labels in a first editing pass — that pass is still an order of magnitude faster than transcribing from scratch.

Where two people speak at once, expect one of them to be lost. That is a physical limit of a single mixed channel, not a model failure.

Upload an audio or video file and the transcript appears on this page. Every paragraph carries the timestamp it came from, so you can find the moment in the original recording without scrubbing.

Speaker labels are available when the recording has more than one voice: each turn is tagged Speaker 1, Speaker 2 and so on, which is what makes interview and meeting transcripts usable rather than a single undivided block.

Files never become public. The upload is processed, the transcript is attached to your job, and nothing is indexed or shared.

Audio: MP3, M4A, WAV, AAC, FLAC, OGG, Opus, WebM, AMR and WMA.

Video: MP4, MOV, WebM and MKV — the audio track is read and the video itself is ignored, so a 2 GB screen recording transcribes the same as the audio ripped from it.

Files up to 500 MB upload directly from your browser to storage, which means a long recording is not limited by any request size cap. A 90-minute MP3 at a normal bitrate is around 80 MB; an hour of video is usually 300–700 MB.

If your file is larger than 500 MB, export audio only from your editor first. It transcribes identically and uploads in a fraction of the time.

Every format is uploaded as-is; converting first adds nothing, and a careless conversion at the wrong sample rate can hurt.

• WAV is about 10 MB a minute in stereo at 44.1 kHz, so an hour can pass 500 MB. Export it as 16 kHz mono, the rate the models use internally, and it drops to about 2 MB a minute with no loss of accuracy.
• FLAC is lossless at roughly half the size of WAV; an hour of stereo is about 300 MB. Worth keeping as a master you can transcribe again later, though it is barely more accurate than a good MP3.
• MOV from an editor or a ProRes export can run to gigabytes. Export audio only (Command-E in QuickTime) and upload that. A screen recording with no audio track gets a clear error, not an empty transcript.
• WebM from a browser recorder often has a missing or zero duration header. The length is measured from the decoded audio, so timestamps come out right anyway.
• OGG and Opus are what WhatsApp, Telegram and Signal voice notes use. They transcribe well even at low bitrates. Share a WhatsApp note to Files rather than forwarding it, so it is not re-encoded.
• AAC streams without an MP4 wrapper decode directly. If a capture was cut off mid-frame, the last second or two may be missing.
• AMR is narrowband phone audio (300 Hz to 3.4 kHz). Ordinary speech comes through; names, spelled-out addresses and digits are where it slips, so check them against the recording.
• WMA usually means an old dictaphone recording. Expect a few points less accuracy than a modern phone recording, and try one file before paying for an archive. DRM-protected WMA cannot be read by any tool.

Ten minutes of audio a day, across up to three files, cost nothing. You see the transcript, the timestamps, the summary, and you can copy all of it.

When a file is longer than your remaining free minutes, the transcript stops at that point and says exactly where it stopped. Nothing is charged and nothing is hidden behind a signup form — you get a real result for the first ten minutes rather than a blurred teaser.

To transcribe the whole file, add credits. Credits are bought once, never expire, and work across every tool on this site. At $0.05 a minute (1 credit), a 45-minute interview costs $2.25 (45 credits) at the $5-pack rate, or about $1.93 from the $12 pack.

A two-person interview transcribed without speaker separation is nearly useless: you cannot tell a question from an answer without listening again.

With speaker labels on, each change of voice starts a new paragraph tagged with a speaker number. Rename them in your own document afterwards — the DOCX download puts speaker names in bold at the start of each turn precisely so a find-and-replace turns Speaker 1 into a real name in two keystrokes.

Diarisation is not magic. Two people talking over each other, a poor phone line, or someone joining halfway through are all cases where labels drift. Fix the three or four places it matters and leave the rest.

On a clear recording with one or two speakers, expect 92–97% word accuracy. That means roughly one wrong word per two lines, usually a name, a unit, or a technical term.

Things that reliably hurt accuracy: recording from across a room, air conditioning or traffic, several people speaking at once, strong accents in a language the model sees less of, and music underneath speech.

Things that help, and cost you nothing: record at arm's length rather than across a table, use the phone's own voice memo app rather than a video call recording where possible, and say proper nouns clearly the first time they come up.

Before you publish or submit anything, check names, numbers and any sentence you intend to quote. The timestamps make that a minute of work rather than a re-listen.

Upload the video, download the SRT, drop it on the timeline. The cues are split on natural sentence boundaries with millisecond timing, which is what Premiere, Final Cut, Resolve, CapCut and every web player expect.

VTT is the same content in the format HTML5 video wants, with speaker names as voice tags so a player can style each speaker differently.

Subtitle files always need a human pass — line lengths, a mis-heard brand name, a laugh transcribed as a word. Editing an SRT that is 95% right takes a few minutes. Typing one from scratch takes an hour per ten minutes of footage.