Why bots know names and local recorders don't
A meeting bot is a participant, so the meeting platform can tell it who is talking. A bot-free recorder hears what your computer hears: your microphone, and the meeting app's audio with every remote participant mixed into one stream. To split that stream it has to group voices by how they sound, which is called speaker diarization. Diarization can say “three different people spoke, and here is when,” but not who they are.
How LokalBot labels speakers
- Two tracks first. Your microphone and the call audio are recorded separately, so your side and the call stay apart before any model runs. By default, speech on the microphone counts as yours (My microphone is me), which also puts your commitments first in action items.
- Voices within each track. After the meeting, Separate voices by speaker, on by default, groups the voices in both tracks on your computer. New Mac installs use Nemotron 3; Windows and Linux (alpha) use pyannote segmentation with TitaNet embeddings.
- Names from you. When you rename a speaker, calendar attendees appear as suggestions. They stay suggestions: even a one-on-one invite doesn't prove who spoke, so LokalBot never applies a calendar name on its own.
- Voices, remembered. When you name a speaker and there's enough clear speech, LokalBot keeps an encrypted voice profile on your computer. That includes people on your calls, who aren't asked. On a Mac, a call voice that closely matches one remembered person is named automatically, and you can undo or correct it; microphone voices are only suggested. On Windows and Linux (alpha), a match is only offered, and Apply Names names the speaker.
These settings and the list of remembered people are in Settings → Meetings → Speaker names.
How accurate is it?
The test ran on a Mac and used four public AMI meeting recordings, each scored twice (headset mix and distant microphone), for 184.55 minutes of audio with four speakers per meeting. Nemotron 3 had a 14.6% diarization error rate with no forgiveness collar, against 43.4% for the previous default, Pyannote Community-1. It found the right number of speakers in all eight cases and ran about 500 times faster than real time. The error rate adds up missed speech, false alarms, and speech credited to the wrong person. See the benchmarks for methods and raw results.
The test has limits: English only, four speakers, and meeting-room recordings rather than audio that went through a video call. Larger calls, other languages, and the speaker models used on Windows and Linux haven't been measured.
When labels go wrong
- Several people on one microphone. In a meeting room, turn off My microphone is me so the room's voices are separated instead of all counted as yours.
- Speakers instead of headphones. Call audio leaks into your microphone and can land on your side of the transcript. Headphones keep the tracks clean.
- Crosstalk and short replies. Overlapping speech and one-word answers are the hardest parts for any diarization model.
- Big calls. Every remote voice shares one track, and more voices mean more chances to merge two people or split one.
When a label is wrong, correct it in the transcript. Your names and corrections stay with the meeting.