Skip to main content
from v3.2.0, miniti can recognise people by the sound of their voice, on the device, so you are “you” from the first sentence and people you have met before are named before they say who they are. it is off by default. it needs speakers set to this mac or this device in settings → transcription, because deepgram’s speaker separation gives nothing to recognise voices from. the first time you turn a voice switch on, a 17 mb speaker model downloads.
a voice fingerprint is biometric data. miniti keeps fingerprints only on the device that made them: 192 numbers per person, never in the cloud copy, never in exports or webhooks, never sent to deepgram, openai or miniti’s servers. forget any of them in settings → transcription; turning remember voices off deletes them all. see the privacy policy.

​
your voice

  1. settings → transcription → voices → turn on your voice
  2. in a meeting, tap your own speaker and choose this is me (on the mac, the microphone speaker is already you)
  3. stop the meeting. the voices section now says how many meetings and minutes of speech it has learned from
from the next meeting on, you are labelled you within about 30 seconds of starting to talk: on the iphone, in a room with several people on one microphone, and in watch recordings. every meeting where you are marked as you teaches it a little more.

​
remember voices

turn on remember voices. when a meeting from google calendar names a speaker and matches them to a guest (see speaker names), their voice is kept under that guest’s email. next time they talk for 10 seconds or more, they get their name, calendar or not. the voices section lists everyone with a bin icon to forget them.

​
when it is not sure

a voice that resembles someone, but not closely enough to be sure, is not applied on its own. tap the speaker: the rename sheet says sounds like sam with a confirm button. confirming names the speaker and teaches the fingerprint.

​
what to expect

measured on recordings of the same four people across three days: a good microphone (headset, a call) recognises everyone with no mistakes. a phone or laptop on the table across the room is harder: most people are still recognised, and when it is unsure it leaves the number rather than guessing. a person recorded on a headset one day and on a room microphone the next is the hardest case; expect the confirm prompt more often there.

​
limits

  • needs apple silicon (any iphone, apple silicon mac) and the on-device speaker model
  • no reading-passage enrolment yet: marking yourself in a meeting is how it learns
  • a person with no calendar email and no pick is not remembered
  • fingerprints do not move between your devices