Why not one model?

A speech recognizer is trained to report what was said. That is not the same job as writing it down. It gives you the "um", the false start, the word you said twice while you thought, and punctuation that ranges from approximate to absent depending on the recognizer.

And you get to choose which one. Dictera ships with four: Apple's system engine, two from Nvidia, one from Alibaba, each with its own language coverage. Asking every one of them to also punctuate consistently across 140+ languages is not a thing you can download. So the punctuation, the fillers and the small repairs are a second pass, done the same way whichever recognizer you picked.

What the second model is allowed to change

The cleanup pass runs a Gemma 4 model locally, and its instructions are a short list rather than a general brief. It may remove fillers, hesitation sounds, false starts and accidental repetitions, in any language. It may fix punctuation, capitalization and spacing by the conventions of the language actually being spoken. It may fix a clear grammatical error, using the smallest change that makes the sentence correct. It may fix an unambiguous misspelling, restore a word you spoke in one language and the recognizer wrote phonetically in another script, and correct a brand or product name where the intended name is obvious.

It may not rephrase, summarize, paraphrase, translate, change the style or tone, or swap a correct word for a synonym. It keeps your line breaks. When it is uncertain, it keeps your text. It runs with greedy decoding, so the same sentence gets the same edit every time.

The part that makes this hard

A dictation is usually a message to somebody else, and very often that somebody is an AI assistant. Which means the text handed to the copy-editor is itself a question or a command, arriving in exactly the slot where a small instruction-tuned model expects to find its own task.

"Could you translate this on Japanese please" should come back as "Could you translate this into Japanese, please?" The failure is that it comes back as the Japanese. The instructions spend a whole paragraph on this one point: a question comes back as a question, a request comes back as a request, corrected rather than carried out.

A prompt is advice, so there is also a check

Between the model returning and the text reaching your document, the result is compared against what went in, and anything no copy-edit could plausibly have produced is thrown away. Three things count: the passage came back in a different writing system, the passage grew far longer than the input, or it collapsed to a fraction of it. Each of those is a model answering, expanding or acknowledging instead of editing. When one fires, your original transcript is what gets inserted.

The check is blunt on purpose. It does not try to catch a rewrite that stayed in the same script — "translate this to German", answered in German. No cheap comparison separates that from a heavy but honest edit, and a check that guessed would start discarding real corrections. That case stays the instructions' job.

Blunt also has to mean careful about mixed language, because people dictate that way constantly. Only a wholesale flip counts as a translation. "я установил спотифай on my new laptop" coming back as "Я установил Spotify on my new laptop" is precisely the repair the model was asked for, even though the text tips toward the other script on a few letters.

What you get when it goes wrong

The raw transcript. If the cleanup model is not downloaded, if generation fails or times out, or if the check refuses the result, the recognizer's words are inserted unchanged. There is no state in which speaking produces nothing because the second model had a bad day.

Which model gets the last word

Neither, as it happens. Your spoken commands are applied first, the cleanup model runs next, and your custom replacements run last, after the model — so a rule you wrote for a colleague's name or a product nobody spells right always wins over the model's view of it. History keeps the raw transcript beside the cleaned one whenever the pass changed anything, so you can see what each model contributed.

And cleanup is a switch. Turn it off in Quick Controls (⌥⌘C) and dictation runs with no language model in the loop at all: the recognizer's words, as it heard them, with spoken punctuation still doing its work. Both models run on this Mac either way, which is the reason the whole dictation pipeline keeps working with the Wi-Fi off — and the reason the list of hosts the app contacts is as short as it is.

The four sections, and what each one is for

Notes come out in a fixed order, and the order is fixed on purpose. Notes whose shape changes from one meeting to the next cannot be skimmed, compared or trusted, because you can never tell whether a missing heading means the meeting settled nothing or the model forgot to look.

The summary is two to six sentences on what the session was about and where it landed. It is not a topic list. The concrete detail belongs here - names, numbers, dates, systems - because that is what people come back for. A summary that says "the team discussed the migration" has thrown the meeting away and kept the calendar entry.

Decisions is one bullet per decision actually made, written as a decision rather than as a discussion that leaned one way. "We ship on the 30th and cut scope on search" is a decision. "The team talked about the date" is not.

Action items is one bullet per task, as who - what, with the deadline where one was given. That is the only section anybody will still be reading next week, so it is written to be lifted straight into a tracker.

Open questions is what was raised and left hanging. It is the section people forget to ask for and then need, because the thing that stalls a project is rarely a decision anybody made.

A section that says "None." is doing work

Decisions and action items are never dropped from a set of notes. If the meeting settled nothing, the heading is still there with "None." under it.

The distinction sounds pedantic until you have read notes with no Decisions heading and had to reopen the transcript to find out which of the two things happened. An absent section is ambiguous; an empty one is a statement. Open questions works the other way round, and is omitted when there were none, because a meeting with nothing outstanding is a good meeting rather than a suspicious one.

Where the owner of an action item comes from

This is the part most write-ups quietly get wrong, so it is worth being precise about what Dictera can and cannot know. Everyone except you arrives on one audio channel, so the transcript has two labels: you, and them. There is no voice fingerprinting and no attempt at one.

Names come from somewhere else entirely: the participants you type into Meeting Setup, or the invitee list read from a calendar invite. Even then, a name is used only where the meeting itself makes clear who is speaking or who is being addressed - somebody says a name, or a person names themselves. An action item whose owner the conversation never established is written Unassigned. That is a rule in the prompt rather than a preference, and it holds even when there is exactly one name on the list and attaching every task to it would look tidier.

If you name some of the people but not all of them, the model is told the list is incomplete. Without that, a model given eighteen names out of twenty treats "not listed" as "not present" and hands every commitment to whichever eighteen it was shown - a mistake that is invisible in the notes and still there weeks later when you go looking for who owes what.

"Next Friday" has to become a date

Meetings are full of relative time. Next Friday, end of the month, in two weeks, after the release. Written down as spoken, an action item stays legible for about a week - which is exactly as long as nobody needs to look it up.

The write-up is told when the session took place, so those references are resolved into actual dates. Where a reference is genuinely too vague to place, the speaker's own words are kept instead of a date being invented. That is the general rule everywhere in the notes: never invent a detail, an owner or a date, and where something is ambiguous, write what is known and mark the gap "(unclear)" rather than resolving it with a guess. A write-up you have to fact-check is worth less than no write-up at all.

What good notes leave out

Half of what makes a write-up readable is subtraction. Greetings, small talk, technical difficulties and scheduling chatter are dropped unless they produced something that belongs in a section - "we lost Ana for ten minutes" matters only if a decision was taken without her.

Sections are ordered by consequence rather than chronology, so the item that matters most is first, not the item that happened first. Nothing is padded to make a section look complete, and nothing is repeated in one section that another already says. Direct quotes appear only where the exact wording carries the meaning: a commitment, a figure, a definition. Everything else is written plainly.

When the recording is not a meeting

Summary, decisions and action items describes a working meeting and collapses on everything else people record. A lecture has no decisions and no owners. Forced into that structure, it manufactures action items out of "read chapter four".

So the structure follows the session. Left on automatic, a recording is read for what it is - meeting, lecture or talk, one-on-one, interview, sales call - and the notes say which shape they used, so a wrong guess is one click from corrected rather than a silent distortion. The headings change with it. A one-on-one gets feedback, commitments and concerns raised. A lecture gets key points, key terms, worked examples and the Q&A. A sales call gets what they need, objections, commercial points and next steps. An interview gets strengths, concerns and the answers worth remembering. The rules about not inventing anything are the same in all of them.

Notes are one rung on a ladder

Notes keep the outcome, and the outcome is a few per cent of a meeting. Every other question about a call - what was actually argued, what someone really said, where in the hour that was - needs the material a summary throws away, which is why Meeting Notes writes four different things from the same recording.

Full minutes keep the meeting: every point, argument and figure, in the order it happened, at roughly a third of the length. The transcript keeps the words, cleaned of disfluencies and mis-hearings, in the language they were spoken and never translated. Chapters keep the shape - a timeline of what was discussed when, with times taken from the transcript rather than estimated. Pick before you press Stop and that is what Stop produces; the other three stay a click away.

When the write-up is not quite right

A write-up is a document, not a reply, so it can be reshaped or edited. Notes and minutes can be cut to the action items alone, reorganised by topic, or turned into a recap email you can send the participants; a lecture offers key terms and a study guide instead. Those rewrites are deliberately not offered on the transcript or the chapters, because a reshaped transcript is no longer a record of anything.

You can also just fix it by hand - a misheard name, a line nobody agreed to - without regenerating the nine-tenths that were right. The model's own version is kept, so Revert still works in a session you reopen next month, and an exported file says that a human edited it.

What the notes still owe you weeks later

The test of a write-up is not how it reads on the day. File a session under the meeting it repeats and every open action item across that series can be gathered into one list. It reads back the notes you already have - no model runs, so it is instant and cannot invent a commitment nobody made - and it tells you when some meetings could not be read instead of quietly showing a short list.

That only works if the individual write-ups were honest. Every Unassigned owner, every "(unclear)", every "None." is what keeps the year-long view from becoming fiction. Notes, minutes, transcript and chapters all export to Markdown with the date, the participants and the model that wrote them, and all of it stays on your Mac - see how the data is handled for what leaves it if you point the assistant at a cloud provider. The rest of what the panel does is on the AI Assistant page, and if you are wondering why none of this appears while the call is still running, that decision has its own reasoning.

Four outputs, one ladder

Meeting Notes shipped able to produce one thing: notes. That is the right default and the wrong only option, because a summary answers a single question well and every other question not at all - the material those answers need is exactly what a summary throws away.

So the four outputs are a fidelity ladder rather than a menu of features. Notes keep the outcome, at a few per cent of the session. Full minutes keep the meeting: every point, argument and figure, in the order it happened, at roughly a third of the transcript's length. The transcript keeps the words, with the speech recognition's damage repaired. Chapters keep the shape - where in the session each subject lives.

Each rung discards less than the one above it. That is the whole distinction, and once you know what the document is for the choice is rarely close.

Which one do I actually want?

The useful test is not what kind of meeting it was. It is what you expect to do with the document after the meeting is cold.

The middle two are the ones people under-use. A colleague who missed a design review does not want your summary of it - a summary tells them where it landed and leaves them unable to reopen anything, because the reasoning that got it there is gone. Minutes are written so someone can read the session instead of listening to it: the points made and by whom where the transcript establishes it, the objections, the figures as given, with decisions and action items marked inline at the moment in the discussion where they happened rather than gathered at the end.

What belongs in a set of notes, section by section, is its own question. This one is about which document to ask for in the first place.

Why the transcript behaves differently from the other three

The transcript is the odd rung, in two ways that surprise people.

It is never translated. Every other output is written about the session, so it follows the language you set for the tab. The transcript is the session. Running it through a second language would quietly replace what people said with a paraphrase of what they said, and a record that has been through two languages is no longer evidence of anything. Each speaker stays in the language they spoke, even where that is two languages in one call.

It is also the one output whose length tracks its input. Everything else compresses, so a long meeting can be handed to the model in big pieces and the answer comes back in a small corner of the window. A stretch of transcript needs as much room to come back as it took to send, which is why the same meeting takes more passes as a transcript than as minutes.

What the clean-up changes is narrow and worth being precise about: punctuation, capitalization and sentence boundaries, which speech recognition supplies almost none of and which are most of what makes a raw transcript unreadable; filler and false starts; obvious mis-hearings where the intended word is unambiguous, names especially. What it never changes is the substance - no turn is summarized, dropped as repetitive or skipped as unimportant - and it never changes the speaker. The instruction is to fix the transcription, not the person: hesitancy and phrasing that are theirs stay theirs. Where the audio is past repair it says "(unclear)" instead of guessing.

Chapters are the output that exists because of the timestamps

Every line in the transcript carries the time it was said. Without that there is nothing for a chapter to point at, and "where did we talk about the contract?" has no answer at all.

A chapter is a subject, not a fixed length: twenty minutes on one thing is one chapter, a rapid run through five items is five short ones. The headings say what a stretch was concretely about - "Pricing for the enterprise tier", not "Discussion". The start and end times are taken from the line markers rather than estimated, and the chapters cover the session end to end with no gaps, opening chat and wrap-up included, briefly.

They are also the cheapest way to make a long recording usable at all: on a two-hour call, the first thing worth asking for, and what tells you which stretch deserves a transcript.

Do I have to decide before I press Stop?

You do not have to, but it is worth doing. The Meeting output setting says what this meeting is for, and stopping the recording produces whatever it is set to. Set a meeting to full minutes and Stop gives you minutes - not notes and then a second click. The button and Stop produce the same thing, so there is one answer to "what do I get from this meeting" rather than two.

Everything else stays one click away on the button's own menu, and choosing from that menu runs immediately rather than only changing a preference: someone opening it wants the document, not a setting. Nothing is lost by picking wrong. Each output lands as its own result in the session, so a meeting can end up holding all four, browsable back and forth, each labelled with what it is, when it was made and whether it covers the whole session.

What each one costs before you ask for it

All four cost model time, which is why none of them runs on its own. A session longer than the model's context window is read part by part, and the menu says how many passes each output would take right now, next to its name, before you commit to it.

That number is not decoration. On the built-in on-device model a nine-pass read is minutes of waiting, and the fix - a model with a larger context window - is one menu away. The pass counts also differ between outputs on the same meeting: the transcript is divided into smaller parts than the others, because of the length-preserving problem above, so a long session takes more of them; notes take one extra pass beyond their parts, because the parts are read and then merged into a single set of notes. Minutes, transcript and chapters are joined rather than merged - each part covers a consecutive stretch and the finished document is chronological anyway - which is also why they degrade gently on a small window. More parts, same document.

If a session is longer than even a part-by-part read can cover, the result says so in its own title: it is labelled as the most recent stretch rather than presented as the whole call. Check that label before forwarding anything from a long meeting.

What if the document is nearly right?

Notes and minutes can be reshaped after the fact - cut to the action items alone, reorganised by topic, turned into a recap email you can send round, or, for a lecture, into key terms or something to revise from.

Those rewrites are deliberately not offered on the transcript or the chapters. Every one of them compresses or reorganises, and a compressed transcript is simply no longer the record of what was said; the chapters have the same problem in miniature, being already the shortest true thing about the session. Both still take typed questions, which is what people actually want from them anyway.

Anything can also be fixed by hand - a misheard name, a line nobody agreed to - without regenerating the nine-tenths that were right. All four export to Markdown with the date, the participants and the model that wrote them, and an exported file says when a human edited it.

The short version

Ask for notes by default. Ask for minutes when someone who was not there needs to follow the argument rather than learn the conclusion. Ask for the transcript when the exact words matter. Ask for chapters when the recording is long enough that the problem is finding things in it.

None of this needs a bot in the call, and none of it leaves your Mac unless you point the tab at a cloud provider - what that changes is set out on the security page, and how a recorded call works from the start is on Meeting Notes. If you are wondering why none of these four appear while the call is still running, that decision has its own reasoning.

What dictation actually takes away

A page of writing is a few thousand keystrokes. Dictated, those words cost one key press, and the rest of the work moves to your voice. That is the whole of the benefit, and it is a large one if typing is what hurts.

It is worth being clear about what stays. Dictation does not move your cursor, click a button, switch apps, or select the sentence you want to fix. Those are still hands, and on a bad day they are the hands that hurt. What follows is about making the part Dictera does own as cheap as possible, and being honest about the rest.

Make the trigger as small as possible

One key, never a chord

Dictera's shipped shortcuts are all single modifier keys, deliberately. Dictation is the right Command key, held. Translating a dictation is the same key, tapped twice. Edit Selection is the right Shift key, held; Correct Selection is a double-tap of the right Option key. Nothing asks you to hold three keys and reach for a letter at the same time, which is the shape of shortcut that hurts.

You can change any of them, and the thing worth keeping when you do is the single-key part. A chord you press two hundred times a day is worse than the typing it replaced.

Hold or tap

The hold mode - push-to-talk - keeps a key down for the length of the recording. It is the tidiest mode, because nothing can be left running, and it is the wrong one if sustained pressure is your problem. A paragraph is thirty seconds of holding a key with one finger.

The double-tap costs two taps at the start and two at the end, with nothing in between. If holding is what aches, switch. Settings ▸ Shortcut takes about a minute, and the three modes and their trade-offs are worth reading once before you pick.

The side that hurts less

Modifier keys come in pairs, and Dictera can listen to the left key, the right key, or either. If one hand is worse than the other, bind dictation to the good side and it will never be asked to do the work. Sticking to the right-hand keys has a second benefit: the left Command key stays free for ⌘C and ⌘V, and an ordinary double-tap during editing cannot start a recording by accident.

One tap for a whole paragraph

Two settings together turn a dictation into a single gesture. Set the shortcut to double-tap, then turn on stop on silence, which is off until you ask for it. Dictera watches the input level while you record; once you have spoken for at least a second and then been quiet for two, it ends the recording and writes the text out.

The result is that you tap twice, talk for as long as you like, and stop. Your hands never come back. For a long email that is the difference between one gesture and about fifteen hundred.

Three dials sit behind it, in Settings ▸ Voice. The one to change first is the silence timeout: two seconds suits someone who speaks in finished sentences, and if you pause to find a word, raise it to three or four. Otherwise the pause ends the recording and the rest of your thought arrives as a second dictation, cleaned up separately, with a paragraph break you did not ask for. If your room is noisy, raise the level that counts as voice rather than the silence, so the background does not hold the recording open forever.

Or no tap at all

The third shortcut mode is an ordinary key combination, and its usefulness here is that Dictera does not care what pressed the key. Anything that sends one works: a spare key on a programmable keyboard, a macro pad under your palm, a foot switch. Bound to a pedal and paired with stop on silence, a dictation costs your hands nothing at all.

There is also a route with no keyboard in it. Dictera's menu bar item has Start Transcription at the top, so a mouse, a trackball or a head pointer can begin a recording the same way the shortcut does. It is slower than a tap, and on the days when it is the only thing that does not hurt, that is fine.

The keys you press after the words land

Return, comma, full stop, the arrow keys taking you back to fix something - the small presses add up, and most of them are avoidable.

• Say the structure. "New line" and "new paragraph" insert the breaks, so you never reach for Return mid-dictation. Voice commands are on by default.
• Do not say the punctuation. You can - "comma", "question mark", "colon" all work - but the on-device cleanup model punctuates what you said anyway, so saying every mark out loud is effort spent for a result you already had.
• "Scratch that" throws the dictation away, so a sentence that came out wrong does not have to be deleted by hand. It only fires when it is the whole of what you said; halfway through a paragraph it is just words.

Fixing text without retyping it

The expensive part of writing with sore hands is rarely the first draft. It is the fifty small corrections afterwards, each one a click, a drag and a dozen keystrokes.

Three shortcuts work on whatever is selected, anywhere on your Mac. Correct Selection fixes grammar, spelling and punctuation in place, in one undoable edit. Edit Selection records a spoken instruction instead - hold the key and say "make it shorter" or "turn this into bullet points" - and the rewrite replaces the selection. Translate Selection shows a translation in a floating panel without touching the document. All three are on the features page in more detail.

The honest caveat is that they all start with a selection, and selecting is hands. ⌘A for a whole field costs one press; ⇧ with the arrow keys costs many. Where you can, dictate into an empty field, correct the whole of it, and move on - that is one selection instead of five.

What this does not fix

Dictera writes. It does not drive the Mac. There is no way to say "click send", scroll a page, or move the cursor two words left, and there is no spoken backspace: dictation appends, and a correction is a selection plus one of the shortcuts above.

The tool for the rest of it is already on your Mac. Voice Control, in System Settings ▸ Accessibility, moves the pointer, clicks, and navigates by voice, and it is a different job from turning speech into good prose. Using one for writing and the other for navigation is a reasonable arrangement, and neither is a substitute for the other.

The setup, in four changes

1. Shortcut mode to double-tap, on the side of the keyboard that hurts less.
2. Stop on silence on, with the silence timeout at three seconds if you pause to think.
3. Voice commands left on, and the habit of saying "new paragraph" rather than reaching for Return.
4. Correct Selection and Edit Selection kept for the fixing, so corrections cost a selection rather than a retype.

None of that is a treatment, and none of it makes a bad day a good one. What it does is move the writing off your hands on the days when that is the difference between answering the email and putting it off. The dictation page covers what happens between the key and the words.

How much faster, really?

Dictera makes this estimate itself. Settings ▸ Statistics compares the seconds you spent recording with how long the same number of words would take to type at 40 words a minute, and shows the difference as time saved. Only the counts are kept - words, sessions, seconds - never the text. The weekly summary that reports it is off until you turn it on.

Forty is a reasonable stand-in for a fluent, non-specialist typist, and most people speak at something like three times it without trying. So the arithmetic comes out flattering, and for a long email or a first draft it is not far wrong. The interesting part is what the number quietly leaves out.

Where the time actually goes

A dictation is not one event. Between the key going down and the text appearing, four things happen in order:

• You speak. This is the part the stopwatch sees, and the only part where speech is obviously ahead.
• The audio becomes a transcript. Two of the four speech engines - Apple's Speech Analyzer and Nvidia's Nemotron - transcribe while you are still talking, so when you stop, most of the work is already done. The other two transcribe the whole recording after you stop. That choice is the single biggest lever on how long the pause feels.
• The transcript becomes writing. Spoken commands are applied, the on-device cleanup model punctuates and removes the fillers, and your custom replacements run last so they always win. Cleanup is the slowest step, and the one you can switch off.
• The text is inserted. Dictera reads what is immediately before your cursor, decides on the spacing and the capital, and pastes. Pasting is near-instant regardless of length; Direct Input, which types the characters one by one instead, is off by default for that reason.

None of those steps is slow on its own. Together they are why a two-word dictation can feel slower than typing two words, even though speech won the part that was measured.

When typing wins

Anything short

The steps above are a fixed cost, and a fixed cost is a tax on short text. "ok, thanks" is faster typed than dictated, and will remain so. The break-even sits somewhere around a sentence, and it moves with your setup: a streaming model with cleanup off pushes it down, a batch model that has to load first pushes it up.

Text that has to be exact

File paths, command flags, variable names, license keys, anything where a stray comma is a bug. Speech recognition is built to produce readable prose, and the cleanup model's whole job is to make what you said read better - which is the opposite of what you want here. Turning cleanup off from Quick Controls gets you the raw words, and it is genuinely useful for a code comment or a commit message, but you will still proofread. Typing thirty exact characters is faster than saying them and checking them.

Sentences you have not finished thinking

Typing lets you stall mid-clause with your hands on the keys. Speech does not. A recording that is running while you think is either a long silence in the middle of your audio, or - with stop on silence turned on - two dictations, cleaned up separately, with a paragraph break where you did not want one. This is the honest reason some people bounce off dictation in the first week. Prose you have already composed goes fast; prose you are composing as you go does not.

Editing what is already there

Dictation appends. There is no spoken backspace, and "scratch that" throws away the recording you are in the middle of rather than the sentence you regret. Fixing existing text is a different move: select it and use Correct Selection for grammar and spelling, or Edit Selection to say the change out loud - "make it shorter", "turn this into bullet points". Both are fast. Neither is dictation, and treating one as the other is how a five-minute email becomes fifteen.

The room you are in

An open-plan desk, a shared office, a train, a call you are already on. No setting fixes this one. Dictation is fastest where you can speak in a normal voice without composing a second message in your head about who can hear you, and there is no version of it that is quick in a room where you would rather whisper.

So when should I use it?

The people who get the most out of dictation are not speaking faster. They have sorted their writing into two piles and stopped thinking about it.

The second habit is smaller: say the structure and let the model do the punctuation. "New paragraph" is worth saying; "comma" usually is not, because the cleanup model adds it anyway. Dictating every mark out loud costs real time and produces text no better than dictating none of them.

Making the gap bigger

Four settings account for most of the difference between a setup that feels quick and one that feels like waiting:

• A streaming speech model, if the languages you speak are covered by one. The transcript is being written while you talk instead of after.
• Keep Model Ready, which is on by default for the speech model and off by default for the cleanup model. Turning the second one on trades memory for a shorter pause on the first dictation after a quiet spell.
• The right recording mode. Hold-to-talk has no end-of-recording decision to make, which suits short bursts; the double-tap suits paragraphs. Both, and stop on silence, are worth ten minutes of trying.
• Custom replacements for the names and jargon you would otherwise correct by hand every time. A word you fix twice a day is worth a rule.

What does not help is dictating faster. Recognition accuracy falls off with speed, and every misheard word is a trip back to the keyboard, which is the expensive kind of time.

The test that matters

Words a minute is the wrong measure for most people, because most writing is not typing-bound. It is bound by the sentence you have not decided on, the reply you have been avoiding, and the wrist that hurts by four o'clock. Dictation is not much help with the first. It is a great deal of help with the other two, and neither shows up in a stopwatch.

So the useful test is not whether the words appeared faster than you could have typed them. It is whether the writing got done. Give it a week of ordinary work, look at the statistics if you like the numbers, and discount them by whatever you spent fixing things. If what is left is an hour, that is an hour. The dictation page covers what happens in that pause between the key and the words.

Where to add one

Open Settings, choose Vocabulary, and use Add Replacement under the Replacements heading. Each rule is two fields: the spoken phrase on the left, what you want written on the right. It saves as you type; there is no Save button.

The examples in that pane are the shape of a good rule: "i phone" becomes iPhone, "git hub" becomes GitHub, "doctor smith" becomes Dr. Smith. Write the phrase the way the transcript spells it, not the way you think you said it. If you are unsure, dictate the word once into a text field and copy what came out.

How the matching works

Three rules cover almost everything you will run into:

• Case does not matter. One rule catches "acme", "Acme" and "ACME".
• Whole words only. A rule for "cat" never rewrites "category", so short phrases are safe to add.
• The longest phrase wins. A rule for "John Smith" runs before a separate rule for "John", so the full name comes out right rather than half-corrected.

Replacements are the last step of a dictation, after the cleanup model has finished. An explicit rule you wrote has the final say on how that word is spelled.

Where the rules stop

Replacements shape dictated prose, and only that. They are not applied to a meeting transcript - for a recorded call, names and terms go in the background field in Meeting Setup instead. They are also skipped when you dictate an instruction to edit a selection, because that sentence is a command for the model rather than text to insert.

The list lives in vocabulary.json on your Mac. It is yours to back up, and a second Mac starts with an empty one, because there is no account to sync it through.

Worth ten minutes once

Add rules for the five or six words you correct most: your product, your team, the clients whose names no transcriber has ever managed. It is the same pane that holds the spoken punctuation commands. The rest of what happens between speaking and typing is on the dictation page, and what is stored where is listed on the security page.

Where the switch is

Open Settings, choose Calendar, and press Connect my calendar. A short explainer comes first, then the macOS permission prompt - access is never requested at first run, and nothing is read until you say yes.

Everything comes from the accounts already set up in the Calendar app, so iCloud, Google, Outlook and Exchange all work without a second sign-in. Access is read-only: Dictera cannot write an event, and no bot ever joins the call.

What happens before a meeting

Before a meeting starts is set to "Ask, then record". A countdown appears, and if nobody answers it starts recording - a meeting you didn't record is gone, one you didn't want is a click to delete. Two pickers under it change that: If I don't answer can be set to "Do nothing", and Give me chooses 15 seconds, 30 seconds or a minute.

That number is also the lead time. Dictera raises the countdown that far before the meeting's own start, so answering it leaves the recording running from the first word rather than from whenever you got to it.

Which entries count as meetings

Calendars to watch is a list of tick boxes - unticking one is how a personal block stops looking like a meeting. Within the calendars you watch, Dictera ignores all-day entries, cancelled ones, anything you declined, anything shorter than five minutes or longer than four hours, and anything with neither guests nor a call link. Titles are never read for keywords, because that only works in one language.

Skip the same recurring meeting twice and Dictera stops asking about that one. What it has learned is listed under Individual meetings, where you can override any series or clear the lot.

Then forget about it

A recording started this way arrives knowing which meeting it belongs to and who was invited, which is what lets the notes name owners on action items. The rest of the setup is on the meeting notes page, and what your calendar contributes to a session is listed on the security page.

What a bot costs

A meeting bot is a participant. It needs an invite or a link, it appears in the list with a name like "Notetaker", and everyone on the call can see it. Some tools announce it out loud. On an internal standup nobody minds. On a client call, somebody asks what it is, and the answer, that the audio of this conversation is being sent to a company none of them have heard of, tends to end the recording.

Then there are the meetings where the question never gets asked because the answer is already no: anything under an NDA, an HR conversation, a legal call, a customer's confidential roadmap, a hospital, a bank. These are the meetings people most need notes from and the ones a bot is least welcome in. A vendor that processes the audio is a data processor, someone has to approve it, and usually nobody does.

How Dictera hears the call instead

Two audio streams, both already on your Mac. Your microphone carries your side, and it is on from the first word, with one click to mute yourself when you would rather not be in the transcript. The other participants arrive through a Core Audio process tap: the same audio your Mac is already playing to your speakers or headphones, captured as it is played. Nothing is injected into the call, no invite is sent, no extra participant appears, and the call keeps playing normally.

Because it is the audio your Mac plays, it works with whatever is playing it: Zoom, Meet, Teams, FaceTime, a call in a browser tab. There is no integration to set up and no permission to ask of the meeting host. Meetings do not need the Screen Recording permission either. Microphone and audio-recording access are all it takes.

One detail that matters on loudspeakers: the other side would otherwise be transcribed twice, once from the call and once as room echo through your microphone. Dictera pauses the mic while the other participants are audible, and stands down when you are on headphones.

What comes out

While the call runs, the transcript builds live, each line stamped with the time it was said, in two labels: you, and them. Everyone else shares one audio channel, so the transcript cannot tell their voices apart, and Dictera does not pretend to. Names come from the participants you listed in Meeting Setup, or from the calendar invite, and are used only where the meeting itself makes clear who is speaking.

When you press Stop, or whenever you press summarize, the write-up is produced. Four kinds, from the same recording:

• Notes. Summary, decisions, action items with owners, open questions. Written automatically on Stop.
• Full minutes. The whole meeting in order, at roughly a third of the length. Not a summary.
• Transcript. The actual words, cleaned of mis-hearings and disfluencies, in the language they were spoken.
• Chapters. What was discussed when, with times taken from the transcript.

Pick which one this meeting is for, and Stop produces it; the other three stay one click away. The notes take the shape of the session, so a lecture is not forced into "decisions" it never made, and where a section is empty it says "None." rather than disappearing. An action item whose owner the meeting never established is written as Unassigned, not guessed. Everything exports to Markdown with the date, the participants and the model that wrote it, and you can ask questions about the meeting afterwards and get the passage that answers them.

What stays on the Mac

With the built-in model, all of it. Transcription is Apple's on-device speech engine and the notes are written by a local model, so a meeting can be recorded, transcribed and written up with the Wi-Fi off. The audio itself is never written to disk. There is no recording to play back, only the transcript and what you generate from it. Sessions live in your Application Support folder, the most recent fifty are kept, and you can delete any of them, or switch history off entirely so that nothing is written at all.

If you point Meeting Notes at a cloud provider instead, because you want a bigger model, the transcript goes to that provider on each pass, and the app tells you so before you switch. The security page lists every host the app can contact, and how to verify the list yourself with a firewall or one Terminal command.

What it does not do

• It does not tell voices apart. Two labels, and names only where the meeting makes them clear.
• It does not write the notes live. The transcript is live; the notes are written on Stop, or when you ask. There is a reason for that.
• It does not detect the language mid-call. Each session runs in one language you choose, so a multilingual call needs a minute of preparation.
• It does not keep the audio. Timestamps point at a moment, not a recording.
• It does not attend meetings for you. It records on your Mac, so you have to be on the call. If you wanted a bot because you were not going to be there, this is not a replacement.

The consent question does not go away

Recording laws vary, and some places require everyone's consent. Dictera captures audio your Mac is already playing, but that changes nothing about your obligations. Tell people you are taking notes, the way you would announce a bot, and check what applies where you and the other participants are. It matters more if you let the calendar start recordings for you, because then the meeting opens with a recording running that you did not press Start on. The difference is not that consent stops mattering. It is that the decision is yours to make and to announce, rather than a participant list's to reveal.

Starting one

Press ⌘⌥H and the assistant panel appears. Choose Meeting Notes, press Start as the call begins, and press Stop when it ends. Or connect your calendar, read-only and from the accounts already set up in Calendar: about a minute before each meeting a countdown asks whether to record it, and the notes are filed under the meeting they belong to. The full walkthrough, and everything the assistant can write, is on the Meeting Notes page.

Most punctuation you do not have to say

Every dictation in Dictera passes through an on-device cleanup model before it lands. That model adds the commas and full stops, removes "um" and "so, yeah", and fixes the capitals, so a plain sentence spoken at a normal pace arrives punctuated. If you learned dictation on older software, the reflex to say "comma" after every clause can go. Most of the time it only gets in the way.

What the cleanup model cannot know is where you wanted a paragraph to end, whether that last clause was a question, or that you have changed your mind about the whole thing. Those are the gaps the commands fill.

The nine commands

The phrases are English, whatever language you are dictating in. They are matched as text after the speech model has finished, not as sounds, so they behave the same with every one of the four speech engines Dictera offers. How reliably a given model hears an English phrase dropped into the middle of a Ukrainian sentence is another matter, and it varies. The structural two, "new line" and "new paragraph", survive best: two clear words with a pause on either side.

Speech models often punctuate the command itself, so "New line." or "comma," arrives wearing its own mark. Dictera expects that and swallows the punctuation on both sides of a command. A space is added after a mark when text follows it, and doubled spaces are collapsed, so "hello comma world" comes out as "hello, world" and not "hello , world".

Where they earn their place

Paragraphs

This is the command you will use most. Left to itself, a cleanup model has no way of knowing that "Thanks for the update" and "I have two questions" belong in different paragraphs, and it will happily produce one dense block. Say "new paragraph" at the break and an email dictated in one breath arrives looking like one you typed.

Questions

Rising intonation does not survive transcription. A model decides between a full stop and a question mark from the words alone, and "you're joining tomorrow" reads as a statement. "You're joining tomorrow question mark" removes the guess.

Lists

There is no "bullet" command. Dictate a list as one line per item, with "new line" between them, then add the dashes by hand, or select the lines afterwards and say "turn this into bullet points", which is one hold of a different key. For a short inline list, "colon" followed by the items works the way it would in speech.

With cleanup off, they are all you have

The cleanup model can be switched off, in Settings or from Quick Controls mid-flow, and there are moments when you want that: dictating into a code comment, a terminal, or a form that will be read by a machine, where you want exactly the words you said and nothing added. With cleanup off, the commands are the only punctuation you get. That is when the full set earns its keep, "period" and "comma" included, and when it is worth remembering that a command swallows the punctuation around it, so the result is still clean. If cleanup is on but fails for some reason, you get the raw transcript rather than nothing, with the commands already applied.

Scratch that, precisely

"Scratch that" is not an undo, and it is worth being exact about what it does. It works only when it is the whole of what you said: press the shortcut, say "scratch that", let go, and nothing is typed. The dictation is thrown away before it reaches the cleanup model.

Its job is cancelling a recording by voice. That matters when your hands are elsewhere, or when you are using stop on silence and there is no key to release and no Esc within reach. It does not remove text already on the page. For that, ⌘Z in the app you are in does what it always did. Said in the middle of a sentence, "scratch that" is just two more words, and they will be typed.

When a command is heard as a word

The failure that does happen is the reverse one: you meant the word. Dictate a paragraph about anatomy and "the colon" gains a stray mark; a note about the end of a sentence loses its "period". Voice commands are a switch in Settings, not a fixture, and the reference list of phrases sits next to it. Turn them off for the session and back on afterwards. The cleanup model still punctuates ordinary sentences either way.

A whole email, one hold of the key

Said, in one recording:

Hi Sam comma thanks for the update new paragraph two questions colon new line does the beta still ship on Friday question mark new line and do you need anything from us before then question mark new paragraph thanks comma Ana

Typed:

Hi Sam, thanks for the update.

Two questions:
Does the beta still ship on Friday?
And do you need anything from us before then?

Thanks, Ana

Half of those commands were optional. The cleanup model would have supplied the comma after "Hi Sam" and the full stops on its own; the paragraph breaks and the question marks are the ones that needed saying. That is the habit worth building: say the structure, and let the model handle the rest.

Everything above works in any app that takes text, because the commands are applied before the text is typed. The dictation page has the rest of what happens between the key going down and the words landing.

What an account is usually for

Three things, in most apps. Identifying a paying customer, so the app knows whether to run. Syncing data between devices, so your history follows you. And knowing who the users are, which is the polite name for marketing. Dictera needs the first, refuses the third, and does without the second on purpose.

What replaces it

A license key. You buy at Lemon Squeezy, who handle the card; we never see the full card details. The key arrives by email, you paste it into Settings, and that is the whole ceremony.

Activating it sends three things to our licensing service: a SHA-256 hash of a hardware identifier, the key itself, and an activation name that defaults to your Mac's name. That last one is worth knowing, because Mac names often contain a person's name. It is there so you can tell your Macs apart, since one license covers 3 of them. The license is checked again roughly once a day, with some randomness in the timing, and it keeps working with no network at all for 14 days.

Your subscription is managed on Lemon Squeezy's page for your order, not in a Dictera account, because there is not one to put it in.

The trial has no signup either

The 14-day trial starts when you first open the app. No email, no card, no "create an account to continue". The app writes the moment of first launch down, in three places so that a reinstall does not reset it, and the licensing service keeps its own verdict against the same hardware hash, so going offline after the trial ends does not restart it either. Every feature is unlocked during the trial, including the assistant.

What you give up

• No sync. Your history, custom replacements and settings live on each Mac and nowhere else. If you work on two, each has its own. There is no cloud copy to restore from, so if your history matters, it is a folder in Application Support that Time Machine already backs up.
• No web dashboard. Word counts and session statistics are computed on your Mac and shown in the app. Nobody, including us, has a graph of your usage.
• No password reset, because there is no password. If you lose the key, it is on your Lemon Squeezy order page.

What you get

• Very little to breach. The licensing service holds what a sale needs: the email the receipt went to, the key, and the hashed identifiers and names of the Macs it is activated on. That is the whole of what a breach could expose, and none of it is anything you said or anyone said to you.
• Nothing to opt out of. No newsletter checkbox, no "help improve the product" telemetry, no analytics SDK in the binary. The app cannot report your usage, because there is no account to report it against.
• A short list of hosts, which you can check yourself. On default settings the app contacts the licensing service, the update feed on this site, and Hugging Face while a model downloads. The security page has the full table and the one-line Terminal command that shows every connection.
• An app that works with the network off. Once the models are downloaded, dictation, cleanup, translation, correction and on-device meeting notes all run with the Wi-Fi off, and the license carries on for 14 days.

Why we chose this

Dictera exists so that the things you say, and the things people say to you in meetings, stay on your Mac. An account would create the one server-side record that could contradict that, and every feature built on top of it, sync, sharing, a web view of your notes, would pull more of your data towards us. Each of those features is useful. Each is also a reason people cannot use a tool like this at work. We would rather ship the version that can go anywhere.

The honest cost is convenience: two Macs mean two histories, and a lost key means a trip to your order page rather than a reset email. We think that is the right trade for a speech tool. If you disagree, everything above is verifiable in about a minute, so at least you know exactly what you are trading. The privacy policy says the same thing at greater length, and the pricing page has the two plans.

Three ways to start

Hold: push-to-talk

Hold a modifier key, ⌘ say, or the right ⌥, and talk. Let go and the words land. The recording is exactly as long as your finger is down, so there is nothing to stop, nothing to cancel, and no way to leave a microphone open by mistake.

It suits short bursts: a chat message, a reply, a command to a terminal, anything under a couple of sentences. It also suits people who dictate in fragments and want each one handled on its own. Where it fails is length. A held key gets tiring after a paragraph, and it occupies a hand you might want for scrolling the document you are dictating about.

Double-tap

Tap the same modifier twice, quickly, and the recording starts and stays on until you end it. This is the mode Dictera ships with, bound to ⌘ ×2 on either side of the keyboard. It suits longer dictation, dictation while you read from something else, and the moments you want to lean back from the keyboard.

You end it with the same double-tap, with Esc, or by going quiet, if stop on silence is on. The cost is that you have to remember it is running. The recording indicator on screen is there for exactly that reason.

A key combination

⌃⌥Space, or any chord you like. It behaves like the double-tap, a press to start and a press to end, but the trigger is an ordinary key press, so it can come from anything that sends one: a spare key on a programmable keyboard, a macro pad, a foot switch. If a hands-free start matters to you, this is the mode to look at.

Left, right or either

Modifier keys come in pairs, and Dictera can listen to one side only. Binding dictation to the right ⌘ leaves the left one for ⌘C and ⌘V as usual, and it stops a quick ⌘-⌘ during ordinary work from starting a recording. The same applies to the hold mode: a right-hand-only key is easy to find without looking.

Three ways to stop

• Release the key, in the hold mode. The natural end.
• Press the shortcut again, in the double-tap and key-combination modes.
• Stop talking, with stop on silence on, in any mode.

And Esc, which is the emergency exit rather than a fourth way. By default Esc ends the recording and throws the audio away, which is what you want when the phone rang mid-sentence. A setting turns that around, so that Esc ends the recording and still transcribes what you said. Pick the one that matches why you usually reach for it.

Stop on silence

Off by default, and worth trying. Turn it on and Dictera watches the input level while you record. Once you have spoken for at least a second and then been quiet for two, it ends the recording and transcribes it. Three dials sit behind that sentence:

• The level that counts as voice. A quiet study and a room with a fan need different settings; the input meter in Settings shows where your room sits.
• The minimum speech before it starts watching, one second by default, so that a cough or a door does not end a recording you have not begun.
• The silence that ends it, two seconds by default.

The defaults suit someone who speaks in sentences and stops when the sentence is over. If you think in pauses, raise the silence to three or four seconds, or the recording will end while you are choosing a word and the rest of the thought becomes a second dictation, cleaned up on its own. If you dictate in a noisy office, raise the level rather than the silence, so that the background does not count as speech and hold the recording open.

Combined with the double-tap, stop on silence is the closest thing to hands-free: tap, talk, and the words appear when you stop. The trade is that a pause to think is now the end of a dictation. Some people love that, because it makes them speak in finished thoughts. Others turn it off after an hour. Both are right.

Which to choose

The small things that change the feel

A few settings sit next to the shortcut and matter more than they look:

• Paste or type. Dictera can insert the text in one paste, or type it character by character for apps that only accept real keystrokes. Paste is instant; typing is what a terminal or a remote desktop needs.
• Sounds. A start and a stop chime, in a few themes, or silence. In the hold mode the start chime confirms the key registered before you begin talking, which is worth having.
• Instant readiness. Keeping the speech model loaded in memory means the first dictation of the day starts without a cold start. It costs memory while Dictera is idle; on a Mac with room to spare it is the better trade.
• Keep the text on the clipboard after it is inserted, if you often paste the same thing twice.

None of these are permanent. The mode that felt right for a week of emails may feel wrong for a day of code review. The setting is one panel away, and the recording indicator tells you which world you are in. If you want the words to come out already structured, the spoken commands work the same in every mode.

Why the assistant is different from dictation

A dictation is short. You speak a sentence or a paragraph, the model works out what language it heard, and the next dictation is a fresh decision. Mixed-language dictation works the same way at a finer grain: each word is recognised and written in its own script. The languages on offer depend on which speech model you downloaded, up to 100+ spoken languages with the largest.

A meeting is an hour of continuous speech from several people, transcribed live on the Mac while the same Mac is drafting answers or waiting to write the notes. The assistant transcribes with Apple's on-device speech engine, set to one language for the session, and the languages on offer are the ones that engine supports on your Mac. Detecting a switch mid-sentence, reliably, in real time, across two speakers on one channel, would cost accuracy on every ordinary meeting to help the occasional bilingual one. So the assistant asks you to choose, and then it is very good at the language you chose.

Where you choose it

Assist and Meeting Notes each have their own language setting, in the panel and in Settings, so the standup you run in English and the customer interviews you run in German keep their own. The transcript is produced in that language, and so is the write-up: notes, minutes and chapters come out in the language of the session. Chat is the exception. It captures no audio and answers in whatever language you type.

Set it before you press Start. It is a setting for the session, not a switch to flip when the conversation changes language.

Assist answers in the same language

The rule covers answers as well as transcripts. Assist drafts its suggested answers in the session's language, so an interview conducted in German gets German suggestions, and a support call in English gets English ones. Each mode also chooses its own model, so a fast local model can serve Assist in one language while a larger one writes up meetings in another. What no mode will do is translate: a transcript is kept in the language it was spoken. If a colleague needs the write-up in another language, export it to Markdown, open the file, select the text, and Translate Selection shows it translated in a floating panel without touching the document.

A meeting in two languages

Choose the language decisions will be stated in. Usually that is the language most of the meeting is held in, but not always: a Spanish team presenting to an English client will spend most of the hour in English and settle the action items among themselves afterwards in Spanish. Which of those you need in the notes decides the setting.

What happens to the other language is worth knowing. The engine hears it as the language you chose, so a German sentence in an English session comes out as English-shaped nonsense, or as nothing. Those lines end up in the transcript as noise and carry little for the notes to work with. A decision made in the other language can be missed. The fix is a habit good multilingual meetings already have: restate the decision in the meeting's language. "So, to confirm in English: we ship on Friday, and Ana owns the release."

Names, terms and the background field

Proper nouns are where a fixed language hurts most. A French product name in an English session, a colleague called Zsófia, an internal codename: any speech engine will do its best, and its best will be wrong. Meeting Setup has a background field for exactly this. Put the product names, the project codenames and the people in it, and the write-up uses the spellings you gave rather than the ones the transcriber invented. If you connect your calendar, the invitee list fills the participants in for you. The setup belongs to each meeting, and a new one starts from your defaults, so the standup's vocabulary is typed once.

When dictation is the better tool

Two situations call for the other pipeline. The first is a one-on-one where you both switch language constantly and neither of you will restate anything. The second is your own notes during a mixed-language call: hold the dictation key, say what you want to remember in whatever language it occurs to you, and it lands in your notes app in the right script, punctuated. Translate as you speak goes one step further: say it in one language and have it typed in another, one of 29. The assistant writes up the meeting; dictation captures what you thought about it. They were built as two pipelines because they are two jobs.

Before a multilingual call

1. Decide which language the notes should be in, and set it for the mode you are using, Assist or Meeting Notes.
2. Fill in Meeting Setup: who is on the call, and every name and term the transcriber is likely to mangle, in the background field.
3. Agree at the start that decisions and action items are stated in the session's language.
4. Keep dictation ready for the notes you take for yourself, in whatever language they come.

If the same recurring meeting alternates languages from week to week, the language is the one thing to check before Start; the participants and the background in Meeting Setup carry over on their own, so the rest of the preparation is already done.

None of this takes more than a minute, and it is the same minute you would spend telling a human note-taker which language to write in. Everything else about how a meeting is captured is on the Meeting Notes page, and why nothing joins the call has its own article.

A document that rewrites itself is not a document

Live notes mean re-summarising the meeting every thirty seconds or so. Each pass reorders the points, rewords the summary, promotes something that just came up and demotes something from earlier. You cannot read a paragraph that changes under your eyes, and you would not trust it if you could, because the version you glanced at four minutes ago has already been rewritten twice.

A meeting's shape is only visible at the end. Decisions land in the last ten minutes. A topic discussed for twenty minutes and then dropped should not be in the notes at all, and a live summariser would have written it up three times before it disappeared. An action item's owner is often settled in the "so who takes this?" at the close. Writing the notes once, from the whole transcript, is how the notes get to say "None." under decisions with confidence, or "Unassigned" beside an item, instead of guessing early and correcting late.

What is live instead

The transcript. Every line, stamped with the time it was said, in two labels, you and them, as the call goes on. If you need to check what was just said, it is there.

And Assist, which is the live mode. When the other side asks a question, a suggested answer appears without being asked; hover any line in the transcript and answer that one; ask for a shorter or a more detailed version in one press. Assist is built for the moments during a call when you need something now. Meeting Notes is built for the write-up. You can run both on the same call, each in its own session.

There is also the summarize button, which is the part people miss. "Not live" means "not continuous". It does not mean "not until the end". Press summarize at any point and you get the notes so far, as a document you can read, at a moment you chose because something worth reading had happened. That is a snapshot on demand, which is a different thing from a document that never settles.

What Stop does

Stop writes the output you chose for this meeting: notes by default, or full minutes, a cleaned transcript, or chapters, with the other three a click away afterwards. A long meeting is longer than most models can read at once, so it is read in passes, the panel says which pass it is on, and if the meeting outruns what one set of notes can cover, the result says it covers the most recent stretch rather than presenting itself as the whole thing. When something has to go, it is the oldest speech, because decisions land at the end.

Some edges are handled so that you do not have to think about them. Stopping twice does not stack a second, identical set of notes; only new speech since the last set counts. A run you cancel halfway keeps nothing and leaves the meeting ready to summarise again. If you pressed Stop too early, Continue listening picks the same session back up. And if the calendar started the recording, the countdown at the end writes the meeting up on its way out when nobody answers, so nothing is lost either way.

Why the end is when the model does its best work

A model reading the whole transcript can classify the session before it writes: a meeting, a lecture, a one-on-one, an interview, a sales call. A lecture has no decisions and no owners, and notes that force "decisions" onto it manufacture action items out of "read chapter four". You cannot tell a lecture from a meeting five minutes in. You can at the end.

The same goes for names. The transcript cannot tell voices apart, so an action item gets an owner only where the meeting made it clear, and that clarity usually arrives late. Written live, the notes would attribute early and wrongly. Written once, they attribute where they can and say Unassigned where they cannot.

The cost, honestly

You cannot glance at a running summary. If you want the notes so far, you press a button. That is one click, and we think it is the right click, because it means the document in front of you was written at a moment you chose, from everything said up to then, and it will not change until you ask again. The rest of what Meeting Notes does, and does not do, is on the Meeting Notes page; why nothing joins the call is the other half of the same design.