Skip to main content

Answer · Speaker labels

Why are the speaker labels wrong in my transcript?

Speaker labels go wrong mostly where people talk over each other, on phone or speakerphone audio, when two voices sound alike, and on short replies like "yeah". The words are usually right; only the label at the change is off. Fix it in the editor: rename each speaker once, move single lines to the right speaker, and merge sentences that got split.

Try it on your own file

Free for 3 files a day, up to 30 minutes each. Speaker labels, timestamps, and exports for Word, SRT and NVivo.

Drop audio or video here

MP3, MP4, M4A, WAV, MOV, WEBM, OGG · up to 5 hours

How speaker labelling works

Speaker labelling (diarization) groups stretches of audio by how the voice sounds, then gives each group a label such as Speaker 1. It does not know who anyone is. It only knows that two stretches sound like the same person, or not.

Here, one speech recognition model focuses on the words and another on telling speakers apart, and the two are combined word by word. That keeps the words accurate even when the speaker boundary is a little off.

The errors you will actually see

A sentence split in two, with the last few words given to the next speaker. This happens when someone starts talking before the other has finished.

A short "mm-hmm" or "right" said over someone else, given its own line. Clean mode drops most of these; verbatim keeps them.

Three speakers merged into two, usually two people of similar age and pitch on the same microphone.

The details

Crosstalk is the main cause

When two people speak at once, there is no clean point where one voice ends. Expect the label to switch a word or two early or late.

Phone and speakerphone audio

Phone lines cut the high and low frequencies that make voices sound different, so two callers can sound more alike than they do in person.

Similar voices

Two speakers of the same gender and age on one microphone are the hardest case. Separate microphones help more than anything you can do afterwards.

Very short turns

A one-word answer gives too little audio to recognise the voice, so it often joins the line before or after.

One recording, one voice per track works best

If your recorder or call app can save each person on a separate track, transcribe each track or use the mixed track with clearer separation.

Step by step

  1. Rename the speakers

    Click a speaker name and type the real name or role, such as Interviewer. Every line with that label changes.

  2. Move single lines

    Where one line belongs to the other person, change the speaker of that line only.

  3. Merge split sentences

    If a sentence was cut in two at a speaker change, merge the second part into the line above.

  4. Listen at the boundaries

    Click the line to play from there. Slow playback to 0.75x where people overlap.

Common mistakes

  • Re-transcribing the whole file to fix a few labels. Fixing them in the editor takes minutes.
  • Renaming Speaker 1 line by line. Rename once and every line changes.
  • Recording a group interview with one phone in the middle of a big table. Put the phone closer to the quieter people.
  • Trusting labels at speaker changes in a quote you will publish. Click the line and listen.

Questions people also ask

Can I tell it how many speakers there are?

No, the number of speakers is detected automatically. If two people were merged into one label, change the speaker of their lines in the editor; the list includes a new, unused speaker.

Why did one person get two different labels?

Usually because their voice changed: they moved away from the microphone, switched to speakerphone, or the room noise changed. Move those lines to one label, or give both labels the same name so exports show one speaker.

Do the renamed speakers carry into exports?

Yes. DOCX, PDF, NVivo and MAXQDA exports use your speaker names on every paragraph.

Will clean mode hide speaker mistakes?

Clean mode removes um, uh, short stutters and short "mm-hmm" or "okay" said over another speaker. That removes many tiny wrong-speaker lines, but it does not change who said a full sentence.

Related questions