Diarization vs. transcription vs. speaker identification
Three things get conflated. Transcription turns audio into words. Diarization separates those words by voice — Speaker 1, Speaker 2, Speaker 3 — without knowing who anyone is. Speaker identification puts real names on those labels, usually by matching against a voice you have already tagged or by pulling names from the meeting invite.
You can have excellent transcription with no diarization at all, and it is the failure mode people notice a week later: the words are all there, but "we agreed to push the date" has no owner. For a one-on-one it barely matters. For a five-person call where the point is who committed to what, an undiarized transcript is close to unusable.
Why bot-free notetakers find this harder
A notetaker bot joins the call as a participant, and the meeting platform hands it a separate audio stream per person. Labelling is nearly free — the platform already did the separation. That is the one genuine advantage of the bot architecture.
A bot-free tool captures the audio flowing through your own machine: your microphone plus everyone else mixed into the system output. Nothing pre-separates the voices, so the tool has to do the work itself, from the mixed signal. This is why several bot-free products quietly ship without it. Granola's desktop app, per its own documentation, labels speech only as "Me" or "Them", with real speaker ID limited to in-person meetings on iPhone — and it is a common enough gap that "does it diarize on desktop?" is worth asking before you commit to any botless tool.
What makes diarization go wrong
Accuracy is measured as diarization error rate — the share of audio assigned to the wrong speaker, missed, or falsely detected. Four things drive it up: overlapping speech, where two people talk at once; similar voices, which get merged into one label; short interjections that are too brief to characterise; and poor audio, where compression and room noise flatten the differences the model relies on.
Two practical consequences. Speaker counts drift — a tool that guesses the number of participants will sometimes invent a sixth or collapse two people into one. And the labels are only as durable as the mapping: if speakers are re-numbered per meeting, "Speaker 2" means someone different every call, which is why identification on top of diarization is what actually makes a transcript archive searchable across months.
How LUCI does speaker diarization
LUCI diarizes on-device on Mac and Windows, from the audio captured on your own machine — no bot in the call and no audio sent to a transcription provider to do the separation. Because the labelled transcript becomes part of a memory that spans meetings rather than a file per call, questions like "what did the CFO push back on across our last three calls?" resolve to one cited answer. It keeps working with the network unplugged, since the separation runs locally like everything else.
Frequently asked questions
What is the difference between diarization and speaker identification?
Diarization separates the audio by voice and gives each one an anonymous label — Speaker 1, Speaker 2. Identification attaches a real name to that label, either from a voice you tagged before or from the meeting participant list. Most tools that advertise "speaker labels" are doing the first; check whether they do the second before you rely on names in a transcript archive.
Does Granola do speaker diarization?
Not on desktop. Per Granola's own documentation, its desktop app labels speech as "Me" or "Them" rather than separating individual participants, with real speaker identification limited to in-person meetings on iPhone. If who-said-what matters on multi-party calls, verify this on your own setup before switching to any bot-free tool.
Can diarization work without a bot joining the meeting?
Yes, but the tool has to do more work. A bot receives a separate audio stream per participant from the meeting platform; a bot-free tool captures everyone mixed together through your machine's system audio and has to separate the voices itself. It is harder, which is why bot-free products differ so much here — several skip it entirely.
Why does diarization get speakers wrong?
The usual causes are overlapping speech, voices that sound alike, interjections too short to characterise, and low-quality audio. Diarization error rate measures how much of the audio is mislabelled, missed, or falsely detected. Better microphones and fewer people talking over each other improve it more than switching tools does.
Does diarization work on in-person conversations?
It is the same problem and generally a harder version of it: one microphone in a room picks up everyone at different distances, with more background noise and more overlap than a call. Expect lower accuracy than a meeting where each participant has their own headset.