Transcripts and speakers

Choose a speech model

Transcripts are made on your Mac by a speech model, so the audio never leaves it. The model is chosen in Settings ▸ Local Models, and dictation uses the same one. The pane is laid out in this order:

1. Recommended for This Mac. Your Mac’s chip and memory, the one model it should use, and a line saying why. Until you pick a model yourself, this is the one MindLake uses.
2. Downloaded. The other models already on your Mac. Click one to use it; the tick marks the model in use. A row’s ⋯ menu has Remove Download.
3. More Models. The rest, each with a Download button.
4. Categories. Apple Speech, Multilingual, English Only, Parakeet and Whisper, for browsing.

The pane’s header has a search field, which matches model names, descriptions and categories, and a ⋯ menu with Reveal Models Folder in Finder and Recheck Downloads, which looks again for models already on your Mac.

The recommendation follows three rules:

The five models:

On an Intel Mac, rows say so where a model would struggle: Slow on this Mac for Parakeet and Whisper Small, and Not suited to this Mac for Whisper Large v3 Turbo. Choosing a model changes new transcripts; to redo an old one with it, use Re-transcribe.

Audio and video files you add to a note as sources are transcribed the same way, speakers included.

Reading a transcript

A recording’s transcript is one of its note’s sources. Click the sources count at the top of the note, such as “2 sources”, or press ⌘., then choose Audio + Transcript.

• Play from any word. Click a word and the recording plays from there, with the current word highlighted and kept in view. Space plays and pauses; ← and → skip 10 seconds.
• One line per turn. Every turn starts on its own line with its speaker’s name in the margin and its start time beneath, even a one-word reply. Each voice has its own colour, and your own turns are shown in a softer one. Older transcripts that put short replies in brackets inside someone else’s turn are rewritten this way when you re-run speakers.
• Editing. It is ordinary text: type to correct it.

Filler sounds

Settings ▸ Transcripts ▸ Trim filler sounds is on by default. It hides “um”, “uh”, “erm” and “hmm” (and longer spellings such as “umm” and “uhh”) while you read and when you copy a transcript, and tidies the commas and capitals around them.

• Nothing is deleted. The fillers stay in the file and playback still follows every word. Turn the setting off and every one is back.
• Only sounds that are never words go. “Like”, “so”, “you know”, “actually” and “basically” always stay, and so do “uh-huh” and “mm-hmm”, which mean yes and no.
• One setting for all transcripts. There is no switch per transcript, and it applies on the Mac only.

The transcript bar

The bar above the transcript holds, from left to right:

The speakers chart

The two-people button at the right of the audio bar folds out a chart of who spoke when. Each speaker gets a row with their initials in their transcript colour, their name, their talk time and their share of the talking, and a lane marking every stretch they spoke, with the playhead running across. Click a lane to jump there. Five speakers show; Show all lists the rest.

No speech, or no transcript

A recording that hasn’t been transcribed shows Not transcribed yet, with Transcribe and Paste, which takes a transcript from the clipboard.

• No speech. If MindLake listened to the whole recording and heard no one, the note says “No speech in this recording” and “MindLake listened to all of it and heard no one speaking.”, with Transcribe Again and Paste. This is saved in the vault beside the recording, so every Mac knows it and none offers to transcribe it again. A file under a second long is settled this way without running the speech model.
• A failed run. The note shows “Couldn’t transcribe:” and the reason, with Try Again.

Speakers

Every transcript is split into turns labelled Speaker 1, Speaker 2 and so on. On a call recorded with system audio your microphone is you by definition, so your turns are labelled “Me”, or the name in Settings ▸ Brain ▸ Your name. In a room, MindLake has to tell the voices apart by sound alone.

Set the number of speakers

If voices are merged or split, tell MindLake how many people spoke, you included. Set it in the recording panel while you record, or afterwards:

1. Click the Speakers item in the transcript bar.
2. Under Expected speakers, choose Auto or 2 to 6 speakers.
3. Choose Re-run Speakers Now… from the same menu, then Re-run Speakers to confirm. It takes about a minute per half hour of audio and replaces any edits you typed into the transcript.

With no number set, the note’s attendees are used as a guide.

Name a speaker

Click a speaker’s label and “Who is this speaker?” opens. The note’s attendees are listed first under In this meeting, then your contacts under Other contacts; or type any name and press Save to rename all their turns. Known as… adds a nickname, such as Abi for Abigail, so it is heard correctly in their meetings. Reset puts the label back. If one person was split across two labels, give both the same name.

Known voices

Naming a speaker also teaches MindLake their voice, so later recordings name them for you. The voices from each note’s last speaker pass are kept on this Mac for your 60 most recent notes, so you can name a speaker days later, after a restart, and MindLake still learns the voice. Re-running speakers keeps the names you gave and learns those voices too.

Settings ▸ Known Voices lists every voice MindLake recognises, with a short ID, how many recordings it was learned from and when it was last heard. Each has Rename and Forget, and Forget every voice… clears the lot: transcripts already written keep their names, and new recordings go back to Speaker 1, Speaker 2. Voices stay on this Mac, are never written into your notes and never sync.

Fixing words

Words the model was unsure of are tinted yellow, the least certain pink; Settings ▸ Audio ▸ Flag words below sets the cut-off (55% by default). Click one to open its fix popover: its confidence, a Correct to… field, your saved fixes for that word, suggested replacements, and Hear, Context and Sentence to play the word, three words either side, or the sentence. Apply corrects it, remembers the fix for future transcripts and dictation, and adds the word to your lexicon.

Settings ▸ Transcripts ▸ Clicking a flagged word decides whether the word plays as the popover opens (Play the word immediately, the default) or only when you press Hear.

The lexicon is your list of names and jargon. New transcripts are checked against it and near misses are corrected; known terms show with a blue wash. Add terms in Settings ▸ Transcripts ▸ Lexicon, or select text, right-click and choose Add to Lexicon. A copy is kept in lexicon.txt in your vault; Reveal lexicon.txt in Finder shows it.

Summaries

Summarise, the sparkle button on the transcript bar, writes the transcript up in the note: a Summary, Topics discussed, Decisions, Action items as tasks, and Open questions. A note still named “Recording ·” gets a real title.

Out of the box, summaries are written by Apple Intelligence on your Mac, which needs macOS 26. Once a local Gemma model is downloaded and chosen, they use that instead, and Settings ▸ AI ▸ Advanced can send them to a cloud provider. If you choose a cloud provider, the transcript goes to it; see Privacy and data. The same AI then adds names and jargon to your lexicon, names speakers it can identify, and may tag the note, or file it if it has no folder.