Speaker diarization, explained.
And what the labels can’t tell you.

1. Start with a recording that suits it

Diarization works best when voices are distinct and people take turns. Use a recording you’re allowed to process: a meeting, an interview, a call you recorded with consent, or a podcast episode. Import the audio or video file, share it into the app, or record in the app (up to 2 hours per recording). Very long recordings can be rejected for speaker detection. If that happens, try a shorter recording.

2. Turn on Speaker labels, then send

In the send options, switch on Speaker labels. It appears for Power subscribers, and other plans see it locked. Choose Transcribe for a turn-by-turn transcript. A speaker-labelled job counts as twice its length toward your plan’s usage, so use it when the conversation has more than one voice.

3. Review, then name the speakers

Read the transcript with the recording. Consecutive segments from the same voice are merged into one turn. The labels are anonymous, so decide who Speaker A and Speaker B are, then copy the text and replace the labels with names in your own document. Check short replies and interruptions closely. They’re the easiest turns to assign to the wrong voice.

Each job returns one output. Here is what speaker labels do in each one:

• Transcribe: one line per turn, starting with a label such as “Speaker A:”.
• Translate: the same labels in the same order, with only the speech translated.
• Summarize: the summary keeps your chosen style, such as Action items, and attributes points to a speaker where it matters. It doesn’t repeat every turn.
• Summarize + Translate: that kind of summary, in the target language.
• If labels can’t be kept: the job fails rather than quietly returning text without them, and you can send it again.

Diarization estimates speaker changes from the audio alone. It can’t see faces or hear who is introduced. These situations need the most checking:

• Overlapping speech: when two people talk at once, words can land in the wrong turn.
• Similar voices: two speakers who sound alike may be merged into one label, or one voice split into two.
• Very short turns: a quick “yes” or “right” is easy to attribute to the wrong person.
• Distance and noise: a voice far from the microphone, or a noisy room, makes changes harder to detect.

Speaker A in one recording isn’t the same person as Speaker A in another. Labels are assigned fresh each time, in the order voices are detected. They are not names, and the app doesn’t recognize or remember anyone’s voice.

Labels apply to new jobs sent with the option on. Results processed earlier don’t gain labels afterwards, so send the recording again if you need them. For interviews where quotes matter, the interview transcription guide covers attribution in more depth. For turning a labelled meeting into tasks, see the meeting transcription guide.