Speaker Diarization

Speaker diarization automatically identifies different speakers in a recording and attributes each portion of the transcript to the person who said it. The result is a transcript that reads like a conversation rather than a continuous block of text.

How it works

After transcription, Murmur runs an offline diarization pass on the audio using FluidAudio, a local diarizer that runs entirely on your device. No audio leaves your Mac during this step.

The diarizer analyzes voice characteristics to detect speaker changes and count how many distinct speakers were present. It then aligns those speaker segments with the transcript text.

Diarization runs after transcription completes, so the total processing time for a meeting is transcription time plus diarization time.

Speaker labels

Speakers are labeled as Speaker 1, Speaker 2, and so on, in the order they first appear in the recording. Murmur does not attempt to identify speakers by name — you can rename them manually in the transcript view after processing.

Each speaker block in the transcript shows:

• The speaker label
• A timestamp for when that turn began
• The transcribed text for that segment

Speaker count detection

Murmur detects the number of speakers automatically. You do not need to specify how many people were in the meeting ahead of time. The detected speaker count is shown in the meeting detail view after processing.

Compatibility

Diarization works with any transcription engine — WhisperKit, Parakeet, and Apple Speech all produce output that the diarizer can process. The quality of the transcript itself still depends on the engine you choose; diarization only adds speaker attribution on top.

Tips for best results

• Clear audio matters most. Background noise and echo make it harder to distinguish speakers. A quiet room with people speaking directly toward the microphone works best.
• Distinct voices help. Speakers with very similar voice characteristics (same gender, similar pitch, similar speech patterns) are more likely to be merged or mis-attributed.
• Avoid cross-talk. When multiple people speak at the same time, the diarizer has to make a judgment call about who is speaking. Minimizing interruptions improves accuracy.
• Longer recordings are fine. Diarization accuracy generally improves with more audio per speaker, as the model has more samples to work from.

Diarization is a best-effort process. For highly accurate speaker attribution in critical recordings, review the output and make corrections manually in the transcript editor.