How to Add Speaker Names to a Transcript

This is for anyone with a recording of more than one person, such as a meeting, an interview or a panel, who wants the transcript to say who said what. You will end up with a transcript where every line carries a real name, exported in a format where the names survive.

What speaker identification actually does

The feature behind speaker names is called diarization. It listens to the recording, works out how many distinct voices there are, and labels each stretch of speech with the voice it belongs to. It does this entirely from the sound of the voices. It has no idea what anyone is called and will not recognise a person from a previous recording. The output is Speaker 1, Speaker 2 and so on, in the order the voices first appear.

It does the slow part, though: separating a forty-minute conversation into turns by hand means listening to all forty minutes. The tool does the grouping; you supply the identity.

Turn it on when you upload, or run it later

The easiest moment is the upload: switch on Identify speakers and the labels arrive with the transcript. If you forget, or you only decide you need them after reading the text, open the transcript and press Identify speakers there. It runs on the existing transcript, so your edits stay, and the same button reads Re-identify speakers afterwards if you want to try again with a different speaker count.

1. Upload the file as usual, or start from a page that matches your recording such as meeting transcription or interview transcription.
2. Switch on Identify speakers.
3. If you know how many people spoke, enter the number. The setting accepts one to twenty. A count helps most when voices are similar; an interview with two people set to two will rarely produce a spurious third speaker, while leaving it unset gives the system room to guess. If you are not sure, leave it blank rather than guessing low, because a count that is too low forces two real people into one label.
4. Choose the language and start.

This option is part of the paid plans and is also available with Minute Pack minutes; the Free plan and the try-without-signing-up box on the landing pages transcribe without it.

Rename the labels

When the transcript comes back, each paragraph starts with a label. Play the first paragraph from Speaker 1, work out who it is, and click the label. Type the name and confirm; every line from that voice across the whole transcript changes at once. Repeat for each speaker. For a two-person interview this is a few clicks, and for a six-person meeting the only slow part is working out who is who, which is where the player helps: click any line to hear it.

A few conventions make the result more useful later. Use the same spelling of a name that appears elsewhere in your documents, so that searching works. Decide whether you want full names, first names or roles (Interviewer, Participant) before you start, especially if you are producing several transcripts for one project. For research where participants are anonymised, replace the generic labels with the participant codes you use in the study.

Fix what diarization got wrong

Diarization is not perfect, and the two common mistakes have easy fixes.

One person was split into two labels. This happens when someone's voice changed, for example because they moved away from the microphone or started laughing, or when the recording is noisy. Use the merge option on the labels to combine them into one. Every line from either label is reassigned, and the text itself does not change.

Two people share one label. This is the reverse, and it is most common with similar voices or when the speaker count was set too low. Find the paragraph where the second person starts, split the paragraph at that point, and relabel the new paragraph. Paragraphs can be split and merged freely in the editor, so you can correct the turn boundaries wherever they landed in the wrong place. If this is happening all through a long recording, it can be quicker to press Re-identify speakers with the right number of speakers than to fix every turn.

When it works and when it struggles

The results are good when each person has a distinct voice, people speak one at a time, and everyone is reasonably close to a microphone. Podcast interviews, one-to-one meetings on headsets and recorded calls where each party has their own line all fall into this group.

It gets harder, in roughly this order, with:

• Crosstalk. When two people talk at once, the overlapping stretch has to go to someone, and the words in it are usually wrong anyway.
• Similar voices. Two people with the same pitch and accent can be merged, particularly if one of them speaks rarely.
• Speakerphones and room microphones. A group around one laptop in a meeting room sounds like a single distant source. Expect labels to be inconsistent, and set the speaker count if you know it.
• Many speakers. A twelve-person workshop or focus group will get labels, but the number of swaps you have to fix grows with the number of people, and brief contributions get lost.
• Phone-line audio. Narrow-band phone recordings strip out much of what distinguishes one voice from another.

Audio repair, the denoise option, is a last resort for speaker separation just as it is for everything else. Most files come out better without it, so try it only after an untreated run has failed.

If you control the recording, the fix is upstream. Ask people to use headsets, record each participant on their own track if the software allows it, and keep the microphone close.

Exports that keep the names

Once the names are right, choose an export that carries them.

• DOCX and TXT put the name at the start of each paragraph, with timestamps on or off as you prefer. These are the formats for minutes, interview write-ups and anything that goes to someone who was not there.
• CSV gives you one row per paragraph with the speaker and the text in their own columns, which is handy for coding research interviews in a spreadsheet.
• SRT and VTT caption files can carry the speaker name on each caption. Leave names on for a panel recording where viewers need to follow who is talking; switch them off for a single presenter.

Alternatives

Zoom and Teams cloud recordings can produce transcripts with names taken from the participant list, which works when each person joined from their own device and fails when a room shares one laptop. Whisper on its own does not separate speakers; local pipelines that add diarization exist but take real setup. For most recordings, turning on one switch and renaming a handful of labels is the least effort for a transcript with real names on it.

Try it on your own recording

Upload a multi-speaker file on the meeting transcription page with Identify speakers switched on, then rename the labels and export to DOCX. Interviews and panels work the same way.