Searching for Dragon on a Mac lands you in 2018

Before anything else, the fact that decides this page. Nuance's own support answer for Dragon Professional for Mac 6 states that the product "is being discontinued effective 10/22/2018" and "will no longer be available for purchase", and that after that date "Nuance will no longer provide updates for the Software" — you can read the answer on their support site. The product they sell today is Dragon Professional, which their page labels "Dragon Professional (Windows)" and describes as "optimized for Windows 11" and "backwards-compatible to Windows 10". macOS does not appear on it.

So this is not really a comparison of two things you can install. It is a page for somebody who typed "Dragon for Mac" into a search box, and the useful shape is: here is what Dragon actually does, here is which parts of it a $19 Mac app replaces, and here is which parts it does not touch at all. The second list is longer than the first, and a page that pretended otherwise would be lying to a reader who is about to find out.

What Dragon advertises, in its own words

Nuance's product page is specific, and the specifics matter because they are the yardstick anybody arriving from Dragon is measuring against:

• "Dictate documents 3 times faster than typing with up to 99% recognition accuracy."
• "Front-end (live speech-to-text) and back-end (transcribed from pre-existing audio files) applications."
• "Custom voice commands to insert standard boilerplate text or signatures."
• "Create time-saving macros to automate multi-step workflows by voice."
• "Transcribe audio files or voice files of another single speaker into text quickly", with an "Auto Transcribe Folder Agent (ATFA) to transcribe batch files of audio recordings."
• Volume licensing and the Nuance Management Center for distributing "disclosures, templates, forms, and legal boilerplate across the organization."

That is four separate products wearing one name: a dictation engine, a batch transcription service, a macro language, and an enterprise deployment system. Coii VoiceInput is the first of those four, on a Mac, and nothing else.

The honest inventory

Nine rows where Dragon is either the only option or the better one. That is not a rhetorical concession — it is the reason a professional dictation market existed around this product for two decades, and if your work needs the right column you should be running Windows, not reading a Mac blog.

What a Dragon user actually loses on a Mac

The people who arrive at this search divide into three groups, and they need three different answers.

The first group used Dragon as a typewriter. They dictated letters, reports and email into a document, corrected by hand afterwards, and never wrote a macro in their life. This is most of them. For this group, the whole of what they used is a key that turns speech into text in the window they are looking at, and a small Mac app does that job now in a way that would have been science fiction in 2018.

The second group lived in the command layer. They said "insert closing paragraph", or had a macro that opened a form, filled three fields and moved to the next one. There is no honest replacement for that here. Coii VoiceInput has no commands, no macros and no scripting. What exists on the Mac for that job is Talon, which controls the machine by voice, and the sensible thing is to read that page rather than to be sold this one.

The third group ran a transcription pipeline — interviews, recorded dictation from a handheld, a folder that filled up overnight. That is a different product again. This app has no way to accept a file: it handles speech as it happens and there is nowhere to put a recording. The Mac apps that do that job are MacWhisper and superwhisper, and both of them are better answers for it than anything on this page.

What replaces the typewriter half

For the first group, here is the whole of it. You hold ⌥ Space, you say the sentence, you let go, and about half a second later the words are in the field you were already typing in. There is no window to switch to, no document to copy out of, and no mode to enter and leave.

Three details are worth more than they sound.

The target is chosen when the key goes down, not when it comes up. If a notification steals focus while you are mid-sentence, or you tab to check something, the text still lands where you started. Anybody who has watched a paragraph of dictation arrive in the wrong window knows why this is a feature and not a footnote.

Your clipboard survives. Where the accessibility route into an application is not available, the text is delivered by a synthetic paste, and what was on the clipboard before is put back afterwards. A dictation app that quietly eats whatever you had copied is a small thing that ruins an afternoon.

Nothing is uploaded. The engine runs on your Mac. Model weights are fetched once from a public model host on first run; after that a dictation involves no network at all, and the app works with Wi-Fi off. There is no account to make and no transcript on anybody's server.

The part Dragon never had

Dragon's own material is about accuracy and speed. The argument this app makes is about order of operations, which is a different axis entirely and the one that survives contact with a real working week.

When you let go of the key, the transcript is written to a local history before the engine is asked anything downstream. Everything that can go wrong — a slow model, an app that was quit, an application that swallows a paste, a window that closed while you were talking — happens after the words already exist somewhere you can get at them. The worst outcome is a copy-and-paste. It is never a sentence you have to reconstruct from memory.

If you dictated professionally for years you will have a story about a sentence that vanished. This is the design that makes that story impossible rather than unlikely, and it is the only claim on this page that no competitor's feature list contains.

The money, with no price to compare against

Nuance's page does not publish a figure; it offers a phone number. That is a normal way to sell to organisations and a frustrating one to shop with, so the only honest column is ours.

Coii VoiceInput is $19 once. Three devices on the licence. Updates included. Nothing renews, there is no tier above it, and the thirty-day trial is the entire app with no card and no account. If it does not fit your work you have lost half an hour, and that is the whole of the risk.

Against a subscription this arithmetic gets loud — Wispr Flow at $15 a month is $180 a year, and Willow is priced in the same territory. Against Dragon it is quieter, because a Windows licence bought once in 2016 was also a one-time purchase and people who liked that model are exactly the people reading this. The point is not that $19 is cheap. It is that the model you liked still exists.

About that 99%

Dragon publishes a number and this app does not, and the reason is worth a paragraph because the number is the thing most people think they are shopping for.

"Up to 99% recognition accuracy" is a vendor measurement, and every vendor measurement in this category is taken under conditions the vendor chose: a particular microphone, a particular room, a particular kind of speech, usually after the software has been trained on the speaker. Change any one of those and the figure moves. Two products quoting 99% are not therefore equivalent, and a product quoting nothing is not therefore worse — it has declined to enter a contest where everybody grades their own paper.

What actually determines whether dictation works for you is narrower and duller: your microphone, your room, how fast you talk, and how many proper nouns your work contains. A built-in laptop microphone in an open-plan office is the single largest variable, and it will do more damage to any product's real accuracy than the difference between any two engines on this page. If you are choosing between dictation tools and you have not yet tried a headset, you are optimising the wrong thing.

This is also why the trial exists in the shape it does. Thirty days, the whole app, no card and no account, because the only accuracy figure that means anything is the one you get on your own hardware saying your own sentences with your own colleagues' surnames in them.

What a real week looks like

The measurable difference between a dictation app that sticks and one that gets uninstalled is not accuracy. It is how many times a day you reach for it, and that number is governed by friction rather than by quality.

Sixty presses a day is an ordinary figure for somebody who has stopped thinking about it: a reply in Mail, three lines in Slack, a comment on a pull request, a paragraph into a document, a note to yourself in Notes. Most of those are under five seconds. At that volume, a half-second delay after key-up is invisible and a two-second one is a tax you notice by Wednesday. At that volume, an app that requires you to open its window first is an app you use four times a day instead of sixty, because the window is the friction and it never stops being the friction.

That is why the target-at-key-down behaviour matters more than it sounds. It is the difference between "I dictate when I am doing a big piece of writing" and "I dictate", and only the second one changes how much you type.

Precisely what leaves the machine

"Nothing is uploaded" is a claim everybody makes, so here it is enumerated rather than asserted.

Model weights are fetched once, on first run, from a public model host. A licence activation goes out when you enter your key. There is a clock check at launch. That is the complete list, and none of it happens during a dictation. No audio and no transcript is sent anywhere at any point, there is no account to attach them to, and the app works with Wi-Fi off — on a plane, on a machine that has never been allowed onto a network, in a building where the guest network is the only network.

The distinction that matters against a hosted service is not the privacy policy but the architecture. A policy is a promise about what a company does with data it receives; this is a design in which the data is never sent. The two are not the same kind of assurance, and only one of them survives a change of ownership.

What this app deliberately does not do

Listed plainly, because a page that only lists strengths is an advertisement:

• No Windows, Linux, iOS or Android. macOS 13 or later and nothing else.
• No voice commands and no control of the machine. It types; it does not navigate, click, or delete a paragraph on request.
• No custom vocabulary you train over time. Unusual surnames, internal product names and specialist terminology are where you will notice. Dragon's vocabulary training was one of its best features and there is no equivalent here.
• No transcription of files, meetings or recordings.
• No account, no sync, and no shared settings between machines. Three devices means three separate installs.
• No real-time translation.

If two of those matter to you, this is the wrong purchase and it is better to find that out from a paragraph than from a refund request.

Moving across, if you are moving

There is nothing to migrate in the ordinary sense — a Windows Dragon profile does not travel to a Mac, and the app that would read it does not exist. What you carry across is the habit, and it transfers in about a day.

The two things to redo by hand are the boilerplate and the corrections list. If you had voice commands that inserted standard paragraphs, rebuild them as text snippets in whatever expansion tool you already use; that keeps them working and keeps them out of a dictation app that has no business storing them. If you had a trained vocabulary of names, accept that the first week will produce some wrong spellings and fix them as you go.

Then change one habit: stop composing in a scratch window. The old shape of dictation was speak-into-the-app, then move the text where it belongs. Here the text starts where it belongs, and the productive week begins when you stop opening a document to talk into.

Who should read a different page

• You are on Windows. Then Dragon is a live product with a feature list this app cannot approach, and none of this applies to you.
• You need commands, macros or hands-free control of the whole machine. Read Talon, and read it properly — it is a serious tool and it is free to download.
• You need files transcribed. A dictation key is not that tool.
• You want to try the free option first. Your Mac already dictates; what Apple Dictation does and where it stops is a fair first stop and costs nothing.

Who should buy this one

• You used Dragon to write prose, not to run macros.
• You are on a Mac and would like the same job done without a Windows machine in the room.
• You want the words on your own disk, with nothing uploaded and no account, and you want that to be how the app works rather than a setting you can forget to enable.
• You are done with subscriptions, and $19 once for three Macs reads correctly to you.

If that is the list, the thirty-day trial answers the question better than any comparison table can. It is the whole app, it needs no card, and a week of real work will tell you within a day.