Skip to main content
from 3.8.0 on mac: when local speaker separation fails during a deepgram recording, miniti tries to restore deepgram speaker separation while recording continues. it waits for actual speaker labels before showing recovery. this does not switch an on-device transcription session to the cloud.
miniti transcribes your meetings in real time using deepgram nova-3 with live interim results. it supports 11 languages, separates speakers by audio source, and can diarize multiple people sharing either the microphone or remote/system audio. from 3.11.0, live transcription has fewer interruptions during background noise. this is available on macos; the iphone and ipad update is waiting for app review.

how it works

there are two audio sources:
  • microphone - one or more people speaking near the recording device. a sole mic speaker in a normal macos online call is labeled “you”; shared-mic recordings use neutral speaker labels until a name or mark as you resolves identity. from v2.7.0, a second mic id needs clear, sustained speech before it counts as another person, and it needs more of it while the meeting looks like an online call, so one person on a headset stays one speaker.
  • system audio - audio playing through your mac (zoom, meet, teams, etc.). remote speakers are labeled “speaker 1”, “speaker 2”, …
on macos, mic and system audio are transcribed on separate channels. people around the mac stay on the mic channel; remote speakers stay on the system channel, even when speech overlaps. each channel can contain several distinct speakers.
acoustic loopback can still put your voice on the system channel if speakers bleed into the room mic path. headphones reduce that.
source-based separation and per-channel diarization keep local and remote voices distinct before any ai insights run.
on ios, there is no system audio capture. ios transcribes and diarizes the microphone only, using neutral speaker labels unless you explicitly mark yourself.

choosing the microphone (macos)

from v2.10.0, settings → recording → microphone (or the mic pill on home) picks which microphone miniti records from. the default, system default, follows macos, which moves the default input to earphones the moment they connect, so a call with airpods in used to record from the airpods mic rather than your desk mic. pick your mic once and miniti keeps using it, with the earphones as output only. a chosen mic that is unplugged falls back to the system default and settings says so.

personal dictionary

add product names, acronyms, and uncommon words in settings → transcription → personal dictionary. miniti boosts those terms during live transcription on both macos and ios. the dictionary sits alongside built-in and calendar terms. total capacity is about 100 terms.

corrections (v2.7.0)

a correction is a pair: what miniti heard, and what it should have written (for example many tea to miniti). you can add one without leaving the meeting:
  • macos - select the word in the live transcript. a correct pill appears at the bottom right of the transcript; click it, or press ⌘⇧D, or right-click the selection. type the correct spelling and save. the same control exists on saved meetings in the transcript editor
  • ios - long-press a transcript line, choose correct a word, tap the misheard word, and type the correction
what happens when you save:
  • earlier mentions in this meeting are rewritten straight away (turn off fix earlier mentions to keep them)
  • the rest of this meeting is corrected as it arrives, including the grey interim text
  • a capitals-only fix works too (v2.8.0): lightdash to Lightdash is a real correction, since the matcher ignores case and keeps the capitals you typed
  • the editor tells you when a correction was not saved (the same word, or the dictionary is full) or when it saved but nothing earlier in this transcript matched, instead of closing quietly (v2.8.0)
  • correcting after you press stop still works: the saved meeting, the minutes file, and any webhook resend carry the fix (v2.8.0)
  • the next recording sends the pair to deepgram as a find-and-replace rule and adds the correct spelling to your dictionary
  • saved meetings keep their insights; only the transcript text changes
corrections live under settings → transcription → personal dictionary where you can review and remove them. they are stored on this device only. the limit is 100 corrections. the matching is whole-word and case-insensitive, and a correction cannot contain a colon.

calendar boost

when a meeting comes from google calendar, the event title and attendee names help transcription recognize people and companies more accurately. non-consumer email domains (for example acme.com) are included when available. see google calendar.

supported languages

set a default in settings or pick per meeting from the language dropdown before recording. transcription, insights, and filler-word detection all adapt to the selected language.

multi-speaker diarization

miniti uses deepgram’s diarization independently on the microphone and system channels. you might see:
  • you - the sole mic speaker in a normal dual-source online call
  • speaker 1 (mic) / speaker 2 (mic) - people sharing the recording device microphone
  • speaker 1 - first remote voice
  • speaker 2 - second remote voice
miniti infers whether a meeting appears remote, in-room, or hybrid from call, calendar, and transcript-source evidence. there is no meeting-type picker, and that inference never changes capture routing or meeting boundaries. from v2.10.0 you can have the who-spoke-when decision made on your device instead: settings → transcription → speakers: this mac (this device on ios; apple silicon, off by default) runs nvidia’s nemotron speaker model on the same audio miniti sends to deepgram and uses its answer for every word. deepgram still does the transcription. see speaker names. diarization itself is source- and voice-based, not name-based. on top of it, miniti runs a separate inference pass that does try to resolve real names from the conversation (and from calendar attendees if connected) - so speaker 1 will often turn into jane once there’s enough signal. see speaker names and google calendar.
from 3.7.0: local speaker separation stays off by default and keeps your saved choice. if it is not ready before recording, deepgram can keep separating speakers for a deepgram session. if it fails during recording, miniti shows speaker separation as unavailable and keeps transcription going 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 local separation for the next recording; model preparation or processing errors still need a retry. with local transcription and separation off, the usual microphone labels remain and no speaker-failure warning appears. your voice and remember voices remain opt-in.

private mode (v2.12.0)

settings → privacy → private mode keeps transcription and speaker separation on the mac and pauses miniti cloud sync. it sets transcription and speakers to this mac, and turns off every feature that needs a model in the cloud: summaries, questions, sales, templates, catch-up, speaker names, jev, playbook and investigations. coaching, history, search, export, calendar and configured crm or webhook exports still work. turn it off and your previous choices come back. greek, polish and swedish cannot be recorded while it is on. a lock shows on the floating recording panel and the home session card while it is on. apple silicon only.
from 3.7.0: managed mode reports active on-device recording duration, including private recordings, even with share diagnostics off. it sends duration, engine, day and device/account identifiers, never audio, transcript, meeting titles, speaker identity or a private-mode flag. local minutes do not use the deepgram allowance. offline reports can arrive later, so totals can have gaps. byok does not send this usage reporting. configured exports remain available in private mode.

transcribe on your mac (v2.12.0)

settings → transcription → transcription: this mac runs nvidia’s nemotron speech model on apple silicon instead of sending audio to deepgram. off by default.
  • audio never leaves your mac for the words. the transcript still goes to the ai service, or your own openai account, for summaries
  • one download, about 590 MB, into miniti’s own folder. settings shows the progress and a button to remove it
  • works in english, spanish, french, german, portuguese and italian. dutch and russian run locally too but made more mistakes in our tests. greek, polish and swedish stay on deepgram, and the setting says so
  • live text lands about a second later than deepgram
  • uses no minutes on the miniti plan, and byok needs no deepgram key for these meetings
  • set speakers to this mac as well to tell remote speakers apart; without it every remote word lands on one speaker
  • if the model is not ready when you press record, miniti refuses to start rather than quietly using deepgram. the fallback setting lets you allow deepgram when a local start is unavailable. a meeting that started on your mac never switches to deepgram
in our tests on recorded meetings it made 2 to 3 more mistakes per 100 words than deepgram on a headset call and fewer on a room microphone.

stopping a recording

from 3.7.0: stopping waits for pending transcript before deciding whether an empty recording should be discarded. on mac, local transcription and enabled local speaker separation have more time to finish their last words and labels. finishing is bounded, so this does not guarantee recovery of audio lost during a connection gap or a failed model.

live transcript view

  • text appears as deepgram returns interim results, then firms up when marked final
  • smooth auto-scrolling keeps the newest content visible
  • selectable text - the live and saved transcript, and insights, are fully selectable. drag to select on macos, long-press to select on ios, then copy any range without exporting the whole meeting
  • search any meeting (including live sessions) - press / to focus the search field on macos

what is sent where

in both managed and byok modes, audio streams directly from your device to deepgram over a websocket. miniti’s backend is never in the audio path. on-device speaker separation (v2.10.0) changes nothing here: the audio flow is the same, and the speaker decisions stay on the device. with transcribe on this mac (v2.12.0) there is no deepgram connection at all for that meeting. the difference is the key:
  • managed mode: the backend issues a short-lived deepgram access token for each session. no deepgram account needed on your side. managed mode also stays connected more reliably on longer meetings and brief gaps.
  • byok mode: the app uses your own deepgram key, stored locally on your device; apple apps use the system keychain.

troubleshooting

  • a meeting said recording but nothing reached the transcript (macos, fixed in v2.10.0) - macos stops the microphone engine when the output device changes, which bluetooth headphones and call apps do as one call ends and the next starts. miniti now restarts the microphone whenever that happens and a watchdog restarts it within 8 seconds if audio stops arriving. on older versions, stop and resume the recording
  • no transcript appears - check microphone and (on macos) system audio permission, or check for error messages on the home screen
  • wrong language - set the language before starting, not during
  • remote audio missing - turn on system audio in the home screen and grant system audio recording permission
  • wrong product names - select the word in the transcript and correct it, or add the term to the personal dictionary in settings
  • local and remote speech overlap on macos - keep both mic and system audio on so each stays on its own channel
see troubleshooting for more.
in 3.8.0, initial recording setup says starting audio or connecting transcription. restarting audio and reconnecting transcription identify an actual interruption. quiet audio alone does not mean something failed.