from 3.8.0 on mac: provider speaker separation can recover after local separation fails during cloud transcription. earlier speaker names remain attached to their original segments; unresolved words stay unknown speaker. a connected socket alone does not mean speaker separation has recovered.
when several people share a microphone in the room, miniti does not guess which one is you. names stay neutral until transcript evidence resolves them or you use mark as you. on a normal online call from a mac, the primary mic speaker stays you even if diarization briefly splits your voice into a second id (v2.7.0).
how it works
- miniti starts a meeting with neutral labels (speaker 1, speaker 2, …) from deepgram diarization, exactly like before
- once there are ~8 final transcript segments, miniti runs a separate, lightweight inference pass to resolve real names only when the transcript or calendar hints are clear enough
- it refreshes every ~40 new segments (with a minimum 2-minute gap) as the conversation gives more signal
- when the meeting ends, miniti runs one more pass over the full transcript for any speaker still without a name (v2.7.0). it runs alongside the final summary and never delays it
- on a 1:1 call, when the calendar event has exactly one other attendee and exactly one remote voice, that voice gets the attendee’s name locally without an ai request (v2.7.0)
- names are overlaid on the live transcript and the speaker legend - the underlying diarization (speaker ids) doesn’t change
- with settings → ai → speaker names: jev selected (v2.10.0, managed accounts), attendee matching runs on typesafe’s jev model instead of gpt; see jev attendee naming
on-device speaker separation
from v2.10.0, settings → transcription → speakers: this mac (this device on ios) runs nvidia’s nemotron speaker model on your mac or iphone to decide who is speaking when, instead of deepgram’s speaker detection. audio still goes to deepgram for the words. the model ships inside the app, so nothing is downloaded, and it needs apple silicon. it is off by default and applies to new recordings. in our tests on recorded meetings it put the right speaker on 6 to 20% more of the talking time than deepgram’s live speaker detection, and it does best in a room with several people on one microphone. if the model cannot load, miniti says so under the toggle and keeps using deepgram’s speaker detection.from 3.7.0: if local separation fails during recording, miniti shows that separation is unavailable and uses unknown speaker for later words. earlier names and your manual corrections are kept. recording and transcription continue where possible. from 3.8.0 on mac, miniti tries to restore deepgram speaker separation during a deepgram recording. recovery is shown only after speaker labels return.a slow session does not disable the prepared model for the next recording. if the model fails to load or stops during a recording, settings shows which one happened and a load model again button. it is available once the recording ends. local separation stays off by default, and your saved choice is preserved. on mac, choosing local transcription with separation off keeps the usual microphone labels and shows no speaker-failure warning.
jev attendee naming
from v2.10.0, managed accounts can turn on settings → ai → speaker names: jev. when google calendar is connected, miniti then asks typesafe’s jev model which attendee each unresolved speaker is, choosing only from the attendee list (or nobody). it returns a choice rather than generated text. the same evidence rules apply: a name is only accepted when someone says it near that speaker’s turns. gpt still handles names that are not on the attendee list, the end-of-meeting pass, and every summary. if the classifier is unavailable, naming falls back to gpt without you noticing. jev is off by default.enable / disable
automatic speaker naming is the speaker names picker in settings → ai on both macos and ios. set it to off if you’d rather keep neutral speaker n labels. manual editing is always available regardless of the toggle.what signals it uses
- the transcript itself - self-introductions (“hey, this is jane”), second-person references (“jane, what do you think?”), and similar cues
- calendar attendees - if you’ve connected google calendar, attendee names are passed as a hint list (“known attendees”). the model is told to prefer these when the conversation matches, not to invent names from them. a name that matches an attendee is accepted when someone says it within a few lines of that speaker talking; other names need the mention right next to it
where names show up
- live transcript - in place of speaker n once resolved
- speaker legend / metrics - the same resolved name
- ai insights - summary, action items, meddpicc, questions all see the resolved names
- webhooks and markdown exports -
speaker_namesmap is included and transcript entries carry the resolved label (see webhooks)
editing a name
if the model gets a name wrong (or misses one), rename the speaker yourself:- macos - click the speaker chip in the speaker legend above the transcript. a popover appears with the name field, plus save / clear / cancel.
- ios - tap the speaker chip in the speaker legend. the same rename view appears as a sheet.
privacy
the transcript is sent through the same insights path used for summaries and questions. no extra service or destination is added. in byok mode, the request goes directly to your openai account. on-device speaker separation adds no destination at all: the model runs locally; resulting speaker labels can appear in insights, cloud copies and exports according to your settings. with settings → ai → speaker names: jev selected (managed accounts only), miniti’s servers forward short transcript excerpts, the attendee names, and the speaker ids to typesafe for the matching step; typesafe returns a choice, never text. choose gpt or off and nothing goes to typesafe. managed gpt speaker naming usesgpt-6-luna. 3.6.0 and older byok use pinned gpt-5-mini-2025-08-07; 3.7.0 moves byok speaker naming to luna too. see ai insights for the full data-flow picture.
limits & caveats
- names are a best guess. the model can get it wrong - especially in short meetings or when multiple attendees share a first name
- microphone and remote speakers can both be named when the transcript supports the name
- ios records the microphone only, so speaker separation depends on how clearly the device captures each person in the room
- calendar attendees are only hints. a calendar match can still be wrong; check it before sharing
related
- transcription - how diarization works, including on-device speaker separation
- google calendar - attendee hints
- ai insights - where the same transcript is used

