Transcribe a recording with speaker labels
To get a transcript that shows who said what, use a tool with speaker labeling (diarization). One-off recording: Nerpa's free web tool - drop the file at nerpa.work/tools/free-transcription, no signup, up to 35 minutes, speaker-labeled text in minutes. Recordings regularly: the Nerpa Mac app - on-device transcription with speaker labels on a supported Mac, plus an editor to fix a mixed label in one click. Local Whisper transcribes well but rarely labels speakers on its own. All three ways compared below, with the honest limits of diarization.
What speaker labeling actually is
A raw transcript is a wall of text: everything everyone said, in order, with no attribution. Speaker labeling - diarization - is the step that splits it by voice. Each turn gets a tag (Speaker 1, Speaker 2), and the wall becomes a dialogue:
With labels: Maya: We should move the launch to Tuesday. Josh: I can have the design ready by Friday, but the copy needs another pass. Maya: Friday works. Dmitri: I'll send the brief today.
Same words - completely different document. Labels are what make a transcript quotable, searchable, and usable as minutes. Diarization runs automatically in modern tools; the human job shrinks to verifying and renaming.
Three ways, compared
| Free web tool | Nerpa Mac app | Local Whisper | |
|---|---|---|---|
| Speaker labels | Yes | Yes, plus an editor to fix and rename | Rarely - usually DIY |
| Setup | None - open the page | Install once, sign in | Real setup - models, CLI or wrapper |
| Limits | 35 min/file, 3 files/day, no signup | Unlimited transcription (supported Mac) | None |
| Where audio goes | Deleted right after transcription | Stays in a local folder on your Mac | Never leaves your machine |
| Export | TXT, SRT, VTT with labels | Editable transcript per person | Plain text, format yourself |
| Best for | One-off files | Regular recordings, meetings | Sensitive material |
The short version: for a single file, the free web tool gives you labeled text in minutes with no account. If recordings are part of your week, the Mac app keeps everything local and adds an editor where a wrong label is a one-click fix. Local Whisper is the zero-upload option - strong text, but you'll label speakers yourself.
How to fix mixed labels
Diarization mistakes are predictable - they cluster in specific places, so a targeted review beats re-reading everything:
- Read the first ten exchanges. If the tool swapped the two voices at the start, it usually keeps them swapped - one systematic fix repairs the whole document.
- Check every interruption. Crosstalk ("yeah, but-", "-that's the point") is where labels break: a half-turn gets attributed to the wrong voice. Fix each with one click.
- Trust the content check: if a line makes no sense for the person it's attributed to, the label is probably wrong, not the words. Diarization errors show up as personality mismatches.
Two caveats to keep expectations honest: same-sounding voices (siblings, similar accents on a conference call) raise the error rate, and a speaker who barely talks may get merged into a neighbor turn. Both are quick fixes in an editor - and rare enough not to spoil the document.
Rename labels to real names
The final step takes one minute and doubles the transcript's value: replace Speaker 1 / Speaker 2 with real names once, at their first appearance. Everything downstream - search, quotes, follow-ups - then works by person, not by number. In the Nerpa app this is built into the workflow: conversations are filed under the person you spoke with, so the label is the person's name from the start and the history accumulates in one place - the same per-client structure that keeps call logs alive.
Transcripts that already know who said what
Nerpa records meetings with no bot joining, transcribes on your Mac with speaker labels, and files every conversation under the person - raw audio stays in a local folder.
Latest stable release. Requires macOS 12+
Honest limits of diarization
What speaker labeling can and cannot do, stated plainly:
- It separates voices, not people. Diarization knows "voice A vs voice B" - it has no idea that voice A is Maya. The names come from you (or from the app if you recorded the call and it knows the participants).
- Four-plus speakers degrade accuracy. Two-three distinct voices work well; a six-person round table will need more manual fixes.
- Phone audio is the worst case. Compressed mono audio merges voices; if the call matters, record your own side locally too.
- It never replaces the quote-check. A correct label plus a garbled word is still a wrong quote - replay anything publishable. The interview transcription guide has the full editing routine.
Frequently asked questions
What is speaker labeling in transcription?
Speaker labeling (also called diarization) is the step that splits a transcript by voice: instead of one wall of text, each paragraph is tagged as Speaker 1, Speaker 2, and so on. It is what turns a raw transcript into a readable dialogue - who said what, in order. Most modern transcription tools combine speech recognition with diarization automatically.
How accurate is speaker labeling?
For clear recordings with two or three distinct voices, modern diarization is usually right for most turns, and the mistakes cluster where people interrupt each other or speak over background noise. No tool is perfect: plan a quick review of the first exchanges and interruptions. Distinct voices and a close microphone improve accuracy more than any setting.
Can I export a transcript with speaker labels to SRT or VTT?
Yes for subtitle formats that carry a speaker per line: SRT and VTT show one subtitle block at a time, so the speaker label becomes the first line of each block. The free web tool on this site exports TXT, SRT, and VTT with speaker labels included. Plain TXT without labels is still the better input for documents and editing.
How do I label speakers manually in a transcript?
If your transcript came back without labels, add them in one pass: play the audio at 1.5x speed, insert the speaker name at every turn change, and use keyboard shortcuts so each insertion is one keystroke. It takes roughly the length of the audio. For future recordings, use a tool with diarization - the free web tool on this site or the Nerpa Mac app label speakers automatically.