Speaker diarization
Speaker diarization is the process of partitioning a recording by speaker, answering "who spoke when." It groups the audio into numbered voices (Speaker 1, Speaker 2) and is harder with noisy or overlapping speech. Reline first splits by source (Me vs Other: your microphone vs the call audio), and a beta option uses diarization to number the voices.
We're aligned on the rollout date — what's the gating risk on the auth flow?
The refresh-token edge case. Once that's covered we can ship to the pilot cohort.
Send me the failure modes before EOD and I'll bake them into the test matrix.
Will do — pairing with the platform team tomorrow morning to align on the schema...
In depth
Speaker diarization is the process of segmenting audio by who is speaking, the “who spoke when” problem, so a transcript can attribute each line to a distinct speaker. It's harder than transcription itself, especially with overlapping talk or many participants, and accuracy varies by tool. Voice-based diarization returns numbered speakers such as Speaker 1 and leaves the names to you or to another source, such as the meeting's participant list. A simpler, more reliable split separates audio by source. Reline does both: speaker separation (Me vs Other: your microphone vs the call audio) separates you from everyone else on the call when the call audio is captured, and a beta option numbers the individual voices.
In practice
Diarization matters once a transcript has more than two people in it: a team standup, a panel, a sales call with three people on the buyer's side, or a workshop recorded on one laptop. It is also where transcripts quietly go wrong. Typical errors are one person split into two speakers after a pause, two similar voices merged into one, short replies given to the wrong person, and crosstalk misassigned. Distance from the microphone, room echo and phone-quality audio make it worse. When you choose a tool, test it on one of your own recordings with the number of people you usually have, and check how fast you can merge two speakers or rename one, because some correction is normal.
How Reline does it
Reline separates speakers in two layers. First it splits by source: your microphone is you (Me) and the call audio is everyone else (Other). That split comes from the audio streams, so it holds even when voices sound alike. On top of it, a beta option tells voices apart and numbers them: Separate speakers for the call side, Shared mic for a room sharing one microphone. The numbers are an estimate and can split or merge people, so you can name a voice or mark two voices as the same person. Reline does not recognize voices: in the transcript, the names come from you.
Open a note and press Start to record the meeting, on a call or in a room.
Press Show transcript. Your microphone and the call audio appear on opposite sides.
On a call, open Who is on the other side beside the record button and choose Separate speakers to number each voice.
In a room sharing one microphone, open Who is on this microphone and choose Shared mic. On a normal call, keep Just me.
Click a label such as Speaker 2 to name it in Name this speaker, or pick a voice marked Same voice to merge them.
Questions
- What is the difference between speaker diarization and speaker recognition?
- Diarization answers who spoke when. It groups speech by voice without knowing anyone in advance and returns labels like Speaker 1 and Speaker 2. Speaker recognition matches a voice to a known person, which needs a stored sample of that person's voice. Most meeting tools diarize, then get names from somewhere else.
- How accurate is speaker diarization?
- It depends on the audio. Accuracy is usually reported as diarization error rate (DER), which adds up missed speech, false alarms and speech given to the wrong speaker. Overlapping talk, many speakers, distant microphones and phone audio push it up, and scores from different test setups are not comparable, so test on your own recordings.
- Does Whisper do speaker diarization?
- Not on its own. The open-source Whisper model returns text with timestamps but no speaker labels. To separate speakers, you pair it with a separate diarization model such as pyannote, which tools like WhisperX combine with Whisper. Some hosted transcription services offer diarization as an option.
Stop taking notes.Start shipping outcomes.
Free forever for individuals. Five minutes to install. Your next meeting writes its own notes.