What speaker identification does
Speaker identification, also called speaker diarization, works out how many people are talking in a recording and which parts belong to whom. The transcript comes back as a conversation rather than a single block:
For anyone who quotes people (researchers, journalists, lawyers, UX teams) this is the difference between a usable transcript and one you have to listen to again.
New in 1.0.2: fewer phantom speakers on long calls
The most common speaker-labelling error is not mixing two people up but inventing a third: on a long call, the same voice drifts enough that the model splits it into “Speaker 3”. Scrieb 1.0.2 switches to a new voice model that tells voices apart better and produces fewer of these made-up extra speakers.
The model is about 30 MB and downloads in the background after the update, so there is nothing to set up. Until it has arrived, speakers are identified the previous way.
Whisper alone does not identify speakers
This surprises many people. OpenAI’s Whisper is excellent at turning speech into text, but it has no concept of who is speaking. Speaker labels need a second model, a diarization step, on top.
Many apps run that step on a server, which means uploading your audio. That is exactly what you cannot do with a confidential interview, a client call or a patient conversation. Scrieb runs both steps, transcription and speaker identification, on your Mac or Windows PC. Nothing leaves the device.
How it works in Scrieb
- Record or import. Record a call with “Record from App”, record the microphone, or drop in MP3, WAV, M4A, MP4, MOV and other files. When importing, choose what the files are (interview, meeting, lecture, voice memo); that sets how many speakers Scrieb looks for.
- Scrieb transcribes and labels speakers automatically, on your CPU or GPU, offline.
- Rename speakers once. Change “Speaker 1” to a name in the Speakers tab and every line of the transcript follows.
- Fix a wrong count. If a two-person interview came out as three, set the number of speakers and click “Identify again”.
- Export with the names intact: Word, PDF, TXT or Markdown, formatted as a dialogue.
On calls where the other person plays through your speakers, Scrieb removes the echo your microphone picks up when the recording is finished, so they are not mistaken for you.
Where it matters most
Research interviews
Separate the interviewer from the participant, and participants from each other in small focus groups. Quotes are attributable without replaying the audio.
Meetings and calls
Know who committed to what. Decisions and action items stay attached to the person who said them.
Legal and compliance
Client conversations, hearings and witness statements need attribution, and they must not be uploaded to a third party.
Journalism
Interviews become quotable dialogue; the Interview summary lists key quotes with who said them.
Getting the best results
- Record close to the speakers. A phone in the middle of a table is better than a laptop microphone across the room.
- Avoid crosstalk. Overlapping speech is the hardest case for any diarization model.
- Use headphones on calls for the cleanest separation between you and the other side.
- Keep the original file. Transcribe the original recording rather than a re-export or a video with music over it.
- Two to six speakers is the sweet spot. Large panels with many similar voices are harder to separate.
Speaker identification: local tools compared
| Tool | Speaker labels | Runs offline | Platforms |
|---|---|---|---|
| Scrieb | ✓ every plan, including free | ✓ | Mac and Windows |
| MacWhisper | Pro only | ✓ | Mac only |
| Buzz | ✓ | ✓ | Mac (Apple Silicon), Windows, Linux |
| Vibe | ✓ | ✓ | Mac, Windows, Linux |
| noScribe | ✓ | ✓ | Mac, Windows, Linux |
| Whisper CLI | ✗ | ✓ | Terminal |
| Cloud services (Otter, TurboScribe …) | ✓ | ✗ upload required | Web |
Full comparisons: best local transcription software and open-source Whisper GUIs.
Pricing
Free plan: 1 hour of recording a day, speaker identification included. Pro: no minute limits, one licence for Mac and Windows.
Try it on your own recordings
Frequently asked questions
- What is speaker identification in transcription?
- Speaker identification, also called speaker diarization, splits a recording by who is talking and labels each part of the transcript by speaker. Instead of one block of text, you get a conversation: Speaker 1 said this, Speaker 2 answered that.
- Does Whisper identify speakers?
- No. OpenAI’s Whisper model only converts speech to text; it has no idea how many people are talking. Speaker labels come from a separate diarization step. Most tools run that step in the cloud. Scrieb runs both transcription and speaker identification locally on your computer.
- Does speaker identification work offline?
- Yes. In Scrieb the whole pipeline runs on your Mac or Windows PC. Nothing is uploaded, and after the models are downloaded no internet connection is needed.
- What changed in Scrieb 1.0.2?
- Speaker identification uses a new voice model that tells voices apart better, with fewer made-up extra speakers on long calls. The model (about 30 MB) downloads in the background after the update; until it has arrived, speakers are identified the previous way.
- How many speakers can Scrieb tell apart?
- Scrieb detects the number of speakers automatically. It works best with two to six people who do not constantly talk over each other, which covers most interviews, meetings and calls. If the count is wrong, set the number of speakers in the Speakers tab and run “Identify again”.
- What if a two-person interview comes out as three speakers?
- Set the number of speakers to 2 in the recording’s Speakers tab and click “Identify again”. When you import files, telling Scrieb what they are (interview, meeting, lecture) also sets how many speakers it looks for.
- How accurate is speaker identification?
- It depends mostly on the audio. Clear recordings with distinct voices and little crosstalk get very good results. Phone-quality audio, heavy background noise or several people speaking at once reduce accuracy for every tool, cloud or local.
- Is speaker identification included in the free plan?
- Yes. Speaker identification is in every plan, including the free plan with 1 hour of recording a day. Pro removes the daily limit. Some competitors reserve speaker labels for a paid tier.
- Can I export the transcript with speaker labels?
- Yes. Speaker names are kept when you export to Word (DOCX), PDF, TXT or Markdown; in Word each speaker’s name is bold, so the document reads like a dialogue.