"Runs on your Mac" settles one question. Three more are still open.

Nearly every dictation app's homepage now says some version of "processes on-device" or "runs locally," and it gets read as though it closes the whole subject of what happens to your voice. It closes exactly one part of it. What happens to your voice is not one question — it is at least four, and a vendor can answer the first one honestly while the other three go unanswered, buried in a settings toggle nobody finds, or answered by a business model that needs the opposite of what "local" implies.

The four questions a privacy claim has to answer

1. Where does the speech actually turn into text — on your machine, or on a server somewhere.
2. Does the app need a network connection at all, for anything, during ordinary use — even one whose core feature is local.
3. What happens to the audio and the transcript afterward — kept where, deleted when, or sent anywhere at all.
4. Is any of it used to train or improve a model — for you specifically, or for every other customer of that vendor.

A page that answers only the first tells you nothing about the other three. What on-device speech recognition actually means covers that first question properly; this page is about the fact that vendors routinely let the answer to it stand in for all four, and readers routinely let them.

Why "local" alone is not the test

Apple Dictation is the cleanest example of why "local" needs a second look rather than a nod. Whether your Mac processes what you say on the device or sends it to a server depends on the hardware and a setting in Keyboard preferences — a fact about your machine, not a promise on a marketing page. What Apple does guarantee, checked against its own privacy documentation today, is narrower and more specific than "local": "the things you dictate are sent to and processed on the server, but will not be stored unless you opt in to Improve Siri and Dictation," and that opt-in is off unless you switch it on yourself. That is a real, checkable promise about question three — retention — and it says nothing at all about question one, because for Dictation specifically the processing path is not fixed.

That is not a criticism of Apple's wording. It is the reason a reader cannot take one true sentence about a dictation app and assume it answers a question the sentence never mentioned. A vendor can be completely honest about retention and silent about processing location, or the reverse, and the resulting page reads as reassuring either way if you are not asking which of the four questions it actually addressed.

The test that cuts through the wording: turn off the network

There is one check that answers questions one and two at once, and it does not require trusting anyone's copy: turn off Wi-Fi entirely — the radio, not "background app refresh" — and dictate through an ordinary working session. An email, a reply, one long paragraph you'd normally just say. If everything works exactly the way it does with a connection, nothing about your voice ever had the option to leave the machine during that session. If something breaks — a first run that needs to fetch a resource, a mode that quietly falls back to a cloud model when the network reappears, a licence check that blocks the hotkey — you have found the boundary the homepage did not state.

Wispr Flow is the honest contrast case, and it is not a hidden one: its own pricing page, checked today, lists Mac, Windows, iOS and Android under one account, and its free tier is explicitly capped at a fixed number of words per week — a cap a server has to be watching your usage to enforce. Running the network-off test against it fails immediately, by design, because dictation there is a request to a server, not a computation your Mac finishes alone. That is not a mark against the product — it is what pays for a hundred-plus languages and a vocabulary that follows you across devices, both real capabilities a fully local app cannot fund the same way. The full comparison works through where that trade helps a given reader and where it does not.

The same page states "Opt out of model training at any time" across its paid tiers. Read plainly, an opt-out framing means the default is on until you act — a different shape of promise than Apple's opt-in, off-by-default one above, even though both sentences use the word "opt." Seven vendors, in their own words, on the training question is where that specific comparison is done properly, rather than repeated here.

What being local does not promise, even when it is true

Processing locally answers question one and stops there on its own. It does not automatically answer question three or four — an app could, in principle, transcribe entirely on-device and still upload the resulting text afterward for some unrelated reason: sync, backup, a support-diagnostics feature turned on by default. Local processing is necessary for the network-off test to pass; it is not sufficient on its own to prove it would. Both have to be checked, and checking retention specifically — what a history entry actually stores, and for how long — is its own question, worked through in what happens to your voice recordings.

Where this app sits on all four questions

Coii VoiceInput has no server behind it at all, which makes three of the four questions unusually simple to answer rather than merely reassuring to answer.

Processing happens on the Mac, always — there is no tier or mode that routes to a cloud model instead, because there is no cloud model to route to. Network dependency is limited to two moments that are not the act of dictating: the model weights are fetched once, on first run, from a public model host, and after that the only network activity at all is licence activation and a clock check at launch. Neither happens while you are dictating, and neither requires a connection that is present in that moment — only that one existed once, earlier. Retention is local: the audio and the transcript live in your own history, on your own disk, until you delete them. Training has no channel to travel through in the first place — with nothing ever sent anywhere, there is no version of "trained on your voice" that could apply to it.

The part that is genuinely a trade, not a boast

Being fully local costs something concrete, and a page arguing for it should say so plainly rather than let the architecture read as strictly better. It is the reason the dictionary here does not learn the way a cloud vocabulary can — you type the entries in yourself, on this device, and they do not follow you to a second Mac automatically. It is also why there is no meeting notetaker built into this product's network model: a feature like that genuinely wants a server. A vendor with infrastructure behind it can offer real things a purely local app cannot, and pretending otherwise would be the same kind of half-truth this page is arguing against.

Why this stops being abstract for some readers

For most dictation — a grocery list, a calendar invite, a reply to a coworker — the difference between these four questions is academic. Any reasonable answer to any of them is fine. It stops being academic the moment what you're dictating is client-privileged, a patient's details, an unreleased product plan, or anything covered by a policy that says the device may not reach a network you don't control. For that reader, "not stored unless you opt in" and "never sent" are not two phrasings of the same guarantee — they are two different risk profiles, and only one of them survives an IT questionnaire that asks whether data can leave the device at all. A locked-down laptop, a SCIF, an aircraft with the connectivity switched off mid-flight are not edge cases dreamed up for a marketing page; they are the ordinary working conditions of a real slice of readers, and for them the network-off test above is not a curiosity, it's the actual acceptance criterion.

The list worth bringing to that conversation

If you're the one who has to answer for a tool in an environment like that, the four questions above turn into four concrete checks rather than four things to take on faith:

Does normal operation ever require an outbound connection? Not "can it work offline in a special mode" — does dictating a sentence, today, on your actual machine, ever open a socket. The network-off test answers this directly and does not require reading a single word of documentation.

What does the update mechanism need, and when does it run? An app that is otherwise silent on the network can still phone home on a schedule to check for a new version. That's a different question from whether a dictation reaches the network, and it's worth checking separately, since "the app never uploads your voice" and "the app never talks to the internet" are not the same claim either.

What happens on first run with no connection at all? Some genuinely local apps still need one download before they work — a model, a licence check — and a reader evaluating a tool for a network-isolated machine needs to know whether that first step can happen somewhere else and be carried over, or whether the device needs a connection at least once no matter what.

Is the retention promise a policy or an architecture? A stated policy — "we don't store this unless you opt in" — is a commitment a vendor is making and can, in principle, change; an architecture where nothing is ever transmitted has no policy to change, because there's nothing on a server for a future policy to reinterpret. Both are legitimate; they are not the same strength of guarantee, and the distinction is usually the actual answer an audit is looking for.

Running the test on whatever you already use

The check does not require this app, or any particular app — it works on anything installed on your Mac right now. Turn off Wi-Fi entirely, then dictate for a real session rather than one test sentence: a reply, a note, a paragraph long enough that a fallback would have time to kick in. If it works exactly as it does online, you have independently confirmed questions one and two without needing to trust either the vendor's marketing page or this one. Dictating without an internet connection covers the mechanics for this app specifically, in case the test is the reason you are trying it, and whether Wispr Flow works offline is the direct answer for the one built to need a connection.

The one-line version

Four separate questions get asked about a dictation app and your voice — where it's processed, whether the app needs a network at all, what happens to it afterward, and whether it trains anything — and a single reassuring sentence on a homepage rarely answers more than one of them. The network-off test is the one check that does not require reading anyone's fine print to get a real answer. Try it for thirty days, with the radio off from the first minute, and see how much of the argument above you needed to take on faith.

The related argument about what a fixed one-time price does to the incentive behind all of this — why a product with no server has less reason to want your data in the first place — is in one-time purchase versus subscription, and the case against letting an app quietly rewrite what you dictated is in dictation that does not rewrite your words.