Speakers

Every transcript tells its speakers apart. Until you name them they are “Speaker 1”, “Speaker 2” and so on, numbered in order of first appearance.

Naming a speaker in one recording

Open the recording, click the speaker’s chip (or the name in a turn), type a name and press Save. The name is used in this recording: in the transcript, in the Speakers column and in the summary the next time it is written. This works with every speech-to-text provider.

Having a voice recognised everywhere

ElevenLabs Scribe recognises the people in your ElevenLabs speaker library. To teach it someone:

  1. Open a recording in which the person speaks and find out which “Speaker N” they are. Play their turns if in doubt.
  2. The app has already cut a voice sample of every unknown speaker: speaker-1.ogg, speaker-2.ogg in the recording’s folder. Show speaker-1.ogg in the speaker’s popover finds it. Listen to it to check that it holds one voice.
  3. In the ElevenLabs web app, open Speech to text, Speakers, Add speaker. Enter a name and a speaker ID, and upload the sample.
  4. Transcripts made from now on label that person’s turns with the speaker ID. Open such a recording, click the chip that shows the ID, type the name to show and press Save. This name is used in every recording, and the existing transcripts are updated at once, without uploading anything and without using credits.

“Recognise speakers from the ElevenLabs speaker library” must be on in Settings.

Recordings made before you added the speaker

They have to be transcribed again, which is paid like a first transcription:

  1. In Recordings, select the recordings in which the person speaks, or press Select all with unidentified speakers.
  2. Press Transcribe again… and confirm. The confirmation shows the hours of audio and about how many credits it costs.

When Scribe recognises a speaker, the others may be renumbered (“Speaker 2” becomes “Speaker 1”); their samples are rewritten to match.

When someone is not recognised

Matching depends on the sample. If a speaker is not recognised in recordings of lower quality, add a sample of similar quality to the library and transcribe those recordings again. If you added the same person to the library twice, give both IDs the same name; their turns are then shown as one person.

The voice samples

A sample is made for every speaker who was not recognised, once per transcript. It is the speaker’s longest uninterrupted stretches joined together, at most 59 seconds, cut from the recording without re-encoding. A speaker with less than 10 seconds of usable speech gets none, because ElevenLabs asks for 10 to 60 seconds. Samples you delete are not made again.

Without ElevenLabs

OpenAI, OpenRouter and Whisper on this Mac have no speaker library: speakers are “Speaker N” per recording and are named per recording. The voice samples are made all the same.