One of them answers you. The other one just types.

People arrive at this search from two different places and they want opposite things. One group wants to talk to an assistant hands-free, and for them the answer is ChatGPT and this page is short: nothing here competes with a model that can answer a question, translate a menu or argue about a recipe. The other group has been typing long prompts all day, noticed that talking is four times faster than typing, and went looking for a way to speak into the box. That is not a smaller version of the same product. It is a different product, and this page is about that difference.

The structural difference, in one paragraph

ChatGPT's voice mode is a conversation with an assistant. You speak, it understands, it replies — and the reply lives in the chat. Apple's App Store listing for the ChatGPT iPhone and iPad app describes it exactly that way: tap the soundwave icon "to have a real-time convo on the go", to "settle a dinner table debate, or practice a new language". That is the shape of the feature and it is a good one.

Dictation is not a conversation. Nothing understands you, nothing replies, and nothing decides what you meant. You hold a key, say a sentence, and the sentence appears in the field your cursor was already in — the reply you were halfway through, the commit message, the ticket, the chat window. There is no interlocutor and no output but your own words.

Both start with a microphone. That is the whole of the overlap, and it is why these two things keep landing on the same search.

Where ChatGPT is better, which is nearly everywhere

This is not a close comparison on capability and pretending otherwise would be silly.

It understands what you said. Ask it to turn a rambling five-sentence thought into three tidy bullets and it will. This app has no idea what your words mean and will type the ramble exactly as you said it.

It answers. Questions, drafts, code, explanations, a language you are learning. None of that exists here in any form.

It is on every platform you own. Phone, tablet, browser, and a desktop build — Homebrew's cask for the Mac app points at OpenAI's own download and lists macOS 13 or later as the floor. Coii VoiceInput is macOS 13 or later and nothing else. No iPhone, no iPad, no Windows, no Linux.

There is a free tier. The App Store listing shows the app as free with in-app purchases; the paid plans it names include ChatGPT Go at $8.00 and ChatGPT Plus at $19.99, checked on 4 September 2026. OpenAI's own pricing page is the place to confirm what the plan you would actually buy costs today, because these change.

If what you want is an assistant you can talk to, stop reading and go and use it. The rest of this page is for the person who wants their own sentences on the screen, unaltered, in an application that is not a chat window.

The copy-and-paste tax

Here is the practical shape of using a chat assistant as a dictation tool, described fairly, because plenty of people are doing exactly this.

You want an email written. You open the assistant, you speak, it produces a draft, you read the draft, you edit the bits where it guessed at your tone, you select it, you copy it, you switch to your mail client, you click into the reply, you paste. Then you fix the paste, because it arrived with the assistant's formatting rather than your mail client's.

Seven steps, two applications, and a piece of writing that started out as yours and came back as somebody else's. For a long, hard piece of writing that trade is often worth it. For the forty short things you write in a working day — the two-line reply, the standup note, the comment on a pull request, the message to your partner about dinner — it is absurd, and everybody who has tried it has quietly stopped.

The alternative is the one in the picture. The chat box is a text field, so a Mac-wide dictation key fills it like any other. You speak the prompt, you read it before you send it — which is a real advantage over a live voice conversation, because a prompt you can see is a prompt you can fix — and the assistant does the part it is good at.

What "it types what you said" means

Verbatim transcription is a strange thing to advertise until you have had the other kind. An assistant, or a dictation product with a model tidying its output, will smooth your sentence: it will fix the grammar, drop the false start, choose a more sensible word than the one you used.

That is often an improvement, and it is sometimes a disaster. If you are a lawyer dictating a clause, a doctor dictating a note, a journalist quoting somebody or a novelist writing dialogue for a character who does not speak in tidy sentences, "smoother" is precisely wrong. You wanted the words you said.

This app writes what you said. The stumble is in there, the repeated word is in there, and the sentence you abandoned halfway through is in there, and you delete them yourself in the two seconds it takes. There is no model between your mouth and the field, so there is nothing to disagree with about tone.

That is also, honestly, a limitation. If what you want is a rambling thought turned into a clean paragraph, an assistant does that and this does not. Aqua Voice and Wispr Flow both sit closer to that end of the category than we do.

The table

Six rows to the assistant, and they are big rows. Ours are narrow and they are about a different job: where the text lands, whether it is exactly what you said, what happens with the network off, and what happens when something breaks.

The failure that only one of these can prevent

Every hosted service has the same bad day, and it is not a criticism of any particular one: something upstream is slow or unreachable, and the thing you said is gone. You said it well, too, and you cannot say it that way again.

The order in that figure is the product's actual order of operations and it is the least copyable thing about it. Let go of the key and the row is on your Mac before anything downstream is consulted. Everything that could go wrong afterwards happens to a sentence that has already been saved.

A conversation cannot work this way and should not be expected to. The whole point of an assistant is that the useful output is the reply, and the reply requires the round trip. But it does mean the two products have different worst days: one loses your connection and gives you nothing, the other loses your connection and does not notice, because it never needed one.

The three ways people try this, and how each ends

Worth walking through, because most readers arriving here have already tried at least one of them.

Talking to the assistant and copying the reply out. Described above. It works, it is slow, and it silently turns your writing into the assistant's writing. People keep it for hard drafts and drop it for everything short.

Talking to the assistant and asking it to transcribe you verbatim. A reasonable instinct, and it half works: you are still in a chat window, still copying the result out, and you have added a step where you have to check whether it followed the instruction this time. You have built a slow, uncertain dictation tool out of a fast, capable assistant.

Dictating into the assistant with a dictation key. This is the one that sticks. The prompt is typed by the key, you read it, you fix the word that came out wrong, you send it. Nothing has been rewritten, nothing was copied between applications, and the assistant is doing the job it is actually good at.

The third is not a compromise between the first two. It is what the two tools are each for, arranged in the order that makes both of them work.

Reading the prompt before you send it

There is an underrated advantage to typing a prompt rather than speaking it into a live conversation, and it is not about privacy or speed.

A prompt you can see is a prompt you can fix. Long instructions — the ones with three constraints, a format requirement and an example — are exactly the prompts where one misheard word changes the answer, and exactly the prompts where you want to reread the constraints before you commit. Speak into a live conversation and the assistant has already started answering the version it heard.

The same argument applies to anything with a name, a number or a piece of jargon in it. Dictated text sitting in a box can be corrected in two seconds. Dictated speech that has already been interpreted has to be corrected by having a second conversation about the first one.

None of this makes a live voice conversation worse at what it is for. Hands-free on a walk, practising a language, thinking out loud with something that answers back — those are its jobs, and a keypress at a keyboard is no substitute for any of them.

Where the words go, and who sees them

Coii VoiceInput processes the audio on your Mac. No recordings and no transcripts are uploaded. The model weights are fetched once from a public model host; after that the network carries a licence activation and a clock check when the app launches, and nothing else. Aeroplane mode changes nothing about how it behaves.

A hosted assistant is the opposite arrangement by necessity: your side of the conversation has to reach a server for there to be a reply at all. That is not a flaw, it is what the product is, and the questions worth asking are about retention and training rather than about transmission. Those answers live in OpenAI's own policies and they change, so read them there rather than trusting a comparison page's summary of them.

The practical rule most people settle on: think about which of the two you are about to speak into before you speak. A client's unreleased numbers, a medical note, an unpublished manuscript, anything under an NDA — those are the ones where "it never left the Mac" is not a preference but a requirement.

A working day with both

That is the configuration that works. The assistant gets the jobs that need thinking — the restructure, the translation, the explanation, the first draft of something hard. The key gets everything else, which is most of the day by volume and almost none of it by difficulty.

Nothing about running both is awkward. They are not competing for a hotkey, they are not competing for the microphone unless you have a live voice conversation open while you press the key, and they are not competing for your money in any meaningful sense: one is a subscription to a capability and the other is a $19 purchase of an input method.

One microphone, two things listening

A small practical note, because it is the only way these two get in each other's way. If a live voice conversation is open and listening, and you then hold the dictation key, two things are using the microphone at once and the result will be strange in a way that looks like a bug in whichever of the two you already distrusted.

End the conversation, or leave it, before you press the key. That is the whole of the conflict. There is no hotkey clash to resolve and nothing to configure, because one of them is a key you hold and the other is a button you tap.

What this app is not

Worth stating plainly, because a page mentioning an AI assistant invites the assumption. Coii VoiceInput has no chat, no summarising, no answers, no translation, no custom vocabulary that learns your jargon and no voice commands. It does not record or summarise meetings — that is a separate Coii product and a separate purchase. It is a key that turns speech into text, on one operating system, and every page in this library that claims more than that is wrong.

A cheap way to decide

You do not need a comparison table for this one, you need a tally.

For three days, every time you are about to write something, write down which of two things you wanted. Did you want your words somewhere — a reply, a message, a note, a ticket, a commit — or did you want an answer to something? Do not overthink the classification; the two feel completely different in the moment.

Almost everybody who does this finds the first column is much longer than the second, and finds that the second column is where the difficult, valuable writing lives. Both facts are useful. The long column tells you a dictation key would be pressed constantly. The short one tells you an assistant subscription is earning its money even though it is used far less often.

Which is why the answer for most people reading this is not one of them. It is that they are priced and shaped for different columns: a monthly capability against a one-off input method, and neither substitutes for the other however hard you push it.

Which to install

ChatGPT if you want an assistant to talk to, you want your words improved rather than transcribed, you are on a phone or a Windows machine, or the thing you actually need is an answer. There is a free tier and it costs nothing to find out.

Coii VoiceInput if you want your own sentences, unedited, in whichever of your applications the cursor happens to be in, with nothing uploaded and the words written down before anything can lose them. $19 once, three Macs, macOS 13 or later, and thirty days of the whole app with no card and no account.

Both, most likely, with the key doing the typing and the assistant doing the thinking. If you want to see how the same argument plays out against tools that are actually dictation products, start with Apple Dictation, which is free and already installed, or superwhisper, which is local like this one. If you write mostly inside documents, Google's voice typing is free and can edit by command. And if your instinct is to build it, a Whisper script is a genuine option with a genuine cost.