A voice editor, or a key that types. Not the same purchase.

Aqua Voice is trying to do something more ambitious than dictation. Its own site describes real-time transcription with grammar refinement, custom instructions, a custom dictionary, screen context awareness and — on the top tier — a Realtime Mode and a "Send it" voice command. The pitch is that you talk and finished text appears, shaped the way you asked for it. That is a genuinely different product from a key that writes down what you said.

It is also a hosted service, and that is the whole of the trade. Coii VoiceInput does the work on your Mac, costs $19 once, and cannot do any of the things in the paragraph above. If the ambitious version is what you want, buy theirs. This page exists so you can tell.

Prices, checked today

From their pricing page on 4 September 2026:

• Starter — free, with 1,000 words.
• Pro — $8 a month billed annually, or $10 billed monthly. Unlimited words, custom instructions, expanded custom dictionary.
• Max — $24 a month billed annually, or $30 monthly. Everything in Pro plus Realtime Mode, the "Send it" voice command, and early access to new features.
• Team — $12 per user per month, annually, adding centralised billing, team-wide settings and enforced Privacy Mode.
• Business — custom, adding SSO/SAML and SCIM, advanced reporting, zero data retention and team-wide dictionaries.

A student discount of 70% is listed on Pro and Max. Their site lists macOS, Windows and iOS, with a downloadable Windows build for Windows 10 and 11.

Coii VoiceInput is $19 once, three devices, updates included, nothing renewing. macOS 13 Ventura or later, Apple Silicon or Intel. Thirty days of the whole app with no card and no account.

Pro at $8 a month annually is $96 a year. Max is $288. Those are not unreasonable prices for a service with a model behind it; they are simply a different kind of purchase, and after a year the gap is five times and after three it is fifteen.

The table

Eight rows to them. This is not a close product comparison — it is a choice between two categories of thing.

Where Aqua Voice wins

It edits, and this app does not. Custom instructions, real-time refinement, a "Send it" command on the Max tier. If what you want is to speak loosely and receive something finished, that capability exists in their product and does not exist in ours in any form. We type what you said. Fixing it is your job and your hands'.

A custom dictionary. Listed on Pro and expanded from the free tier. Anyone who dictates the same handful of client names, drug names or internal product names all day knows exactly what that is worth. This app has no way to learn a word, so the same unusual surname comes out the same wrong way indefinitely.

Forty-nine languages. Their number, from their site. Ours is several.

iOS and Windows as well as the Mac. Dictation is a habit and a habit that stops at the edge of one device is a habit with a hole in it. We are macOS and nothing else.

Team infrastructure. Centralised billing, team-wide settings, enforced Privacy Mode, and SSO/SAML with SCIM and zero data retention on Business. If you are buying for thirty people and your security team needs a console, we do not have one, and no architectural argument substitutes for the console they asked for.

Accuracy claims backed by a number. Their site quotes 97.3% on a named benchmark. We publish no such figure. Treat every vendor's benchmark with the scepticism it deserves, including the absence of one.

Where Coii VoiceInput wins

The network is not in the loop

This is the row that is not a feature difference but a shape difference. Aqua Voice is a hosted service; their site offers a Privacy Mode in which nothing is stored on their servers, and zero data retention on the Business tier. Those are meaningful commitments about retention. They are not, and cannot be, commitments about transmission — a hosted transcription service has to receive the audio to transcribe it.

Coii VoiceInput fetches its model weights once from a public model host and is then finished with the network. No audio, no transcripts, nothing uploaded. It works on a plane, on a train through a tunnel, on a laptop that is not permitted to talk to the internet, and in the hour your building's connection is down.

That is not a criticism of a hosted design. Real-time refinement and a model that follows written instructions are things a service can offer and a local utility cannot. It is a trade, and the point of stating it this plainly is that it is the one thing neither vendor can change with a setting.

The words are saved before the engine is asked anything

Release the key and the row is in your history before the engine is consulted. That ordering is the product's actual argument, and it happens to matter most in exactly the situation a hosted service is most exposed to: a request that takes too long, or one that never comes back.

To be fair to them, a good hosted client buffers audio and retries, and Aqua Voice may well do so — their site does not describe it either way, so this page will not pretend to know. What it can say is what ours does, and that you can verify it: dictate a long sentence, quit before the text lands, and go look.

Where the text goes was decided when the key went down

The destination is fixed at key-down rather than key-up. Click into another window mid-sentence — which everybody does while thinking — and the words still go where you started. Where the accessibility API cannot place them, the fallback is a synthetic paste with your clipboard restored afterwards, so nothing you had copied is quietly thrown away.

$19 once

$96 a year for Pro, $288 for Max, against $19 that does not recur. Both are honest prices for what they are; a service with a model behind it has costs every month and should charge every month. But if you are going to use it for three years, the arithmetic is the arithmetic, and it is worth doing before you subscribe rather than after.

Latency, and the thing a table cannot show

The interesting technical difference between a hosted dictation service and a local one is not average speed. It is variance.

A service running on somebody else's hardware has a fast, consistent floor, because that hardware is fast and it is the same hardware every time. What it adds is a round trip, and a round trip is only as reliable as the worst network you use it on. On office fibre it is invisible. On hotel wifi, on a train, in the corner of a building where the signal drops to one bar, it is the entire experience — and it degrades in the least helpful way, which is unpredictably. A tool that is quick nine times and takes six seconds on the tenth trains you to stop trusting it.

A local app has the opposite profile. Its floor is set by your machine, so an older Intel Mac is slower than a new one and a long sentence takes longer than a short one, and no amount of good engineering changes that. What it does not have is variance from anything outside the machine. The half-second is the half-second in a basement, on a plane, and during the twenty minutes your building's connection is out.

Which of those you prefer is genuinely a matter of where you work. Somebody at one desk with good internet has very little reason to care. Somebody who dictates in a car park between site visits has an enormous one.

The other half of responsiveness is not latency at all: it is whether the app is listening the instant the key goes down. If there is a warm-up, your first word disappears, and after that happens a few times you begin pausing before you speak, which is slower than typing and feels worse. That property is worth more attention than any published millisecond figure, and it is testable in about a minute in either product.

The cost of an account

A small structural difference that turns out to matter more than it reads.

Aqua Voice's paid tiers require an account, which is normal and necessary for a hosted service — there has to be something to attach a subscription to. That account is also where the team features, the shared dictionaries and the enforced Privacy Mode live, so for an organisation it is a feature rather than overhead.

Coii VoiceInput has no account. There is a licence key, three devices, and a launch-time check. Nothing signs in, nothing syncs, and there is no profile anywhere with your name on it. The consequences run both ways: nothing to breach and nothing to leave behind, but also no way to move settings between your machines, no shared vocabulary across a team, and no administrator view of anything.

For one person on one Mac that trade is almost entirely in our favour. For fifteen people in a company with a security review, it is almost entirely in theirs, and pretending otherwise on a page like this would be the fastest way to lose a reader who has actually run a procurement process.

What the history holds

The cost of that design is honest and should be said: there is no sync. What you dictated on the laptop is not on the desktop, and there is nothing on a phone at all. A hosted product gets that for free and we do not get it at all, because the mechanism that would provide it is the mechanism this app exists to avoid.

Who is actually choosing between these two

In practice the decision usually comes down to one of three situations.

You want the finished text. You dislike editing, you would rather describe what you mean than phrase it, and a tool that tidies as you go sounds like the whole point. That is Aqua Voice's design. Nothing here competes with it and this page should not try.

You want the words exactly as you said them. You are a writer, or a developer writing commit messages, or somebody dictating notes that will be read back as evidence of what was said. Refinement is not a feature in that context, it is interference, and a tool that improves your sentences is a tool you have to check. This app writes down what you said.

You cannot send the audio anywhere. A locked-down machine, a client agreement, work that simply is not going through a third party, or a connection that is not reliable enough to depend on. That is not a preference and no privacy mode resolves it, because the constraint is on transmission rather than storage.

If you are in the third group, stop reading comparisons and filter the whole category by that one property first. It removes most of the market and makes the rest of the decision much smaller.

What neither of them does

Neither app is a general voice control for your Mac. Aqua Voice's "Send it" command is a specific action inside its own flow, not a way to switch applications or delete a paragraph by speaking. Full computer control by voice is Talon's territory and Talon does it properly.

Neither is a transcription service for recordings you already have; both are for speech as it happens. And neither replaces reading your own work — punctuation comes out approximately right and proper nouns approximately wrong in every product here, whatever the benchmark says.

The week that decides it

Both give you a way to find out cheaply: their free tier is 1,000 words and ours is thirty days of the entire app with no card. Use both.

The mistake to avoid is testing in a text editor with sentences you have composed in advance. Everything passes that. Test by dictating into the three applications you actually spend your day in, at the speed you actually speak, with whatever noise is normally in the room, and pay attention to three things: how often a name comes out wrong, which app the text does not reach, and how you feel about the version of your sentence that arrives. That last one is the whole difference between these two products and it is not measurable.

Run them a week apart rather than together. Two dictation tools listening for hotkeys is a bad experiment, and on the same key it is a broken one.

What switching costs

Very little, in either direction, and it is worth saying because switching anxiety keeps people on tools they have already stopped liking.

Dictation accumulates no library. The output went into other people's applications as you spoke, and it is now just text in Slack, in Mail, in a document. There is no archive to migrate, no proprietary format, and no project file only one of these can open. Whatever history each app keeps is a convenience rather than an asset that traps you.

If you are leaving a subscription, do two things properly. Export whatever the service holds for you while the account is still active, because access usually ends when the billing does. And cancel from the vendor's own account page rather than assuming that deleting the app stops the charge — it does not, in any product, and that is the single most common way people end up paying for a tool they stopped using in March.

Then give the new key a full working week before you judge it. The hardest part of changing dictation tools is not accuracy, it is the reflex: a different key in a different place under a different finger. Almost everybody who abandons a switch abandons it in the first three days, while the old habit is still winning, and what they report afterwards is that the new one was worse. Usually what was worse was the third day.

Which one to buy

Buy Aqua Voice if you want the text improved rather than transcribed; if a custom dictionary and custom instructions are things you would genuinely configure; if you dictate on iOS or Windows as well as a Mac; if you write in many languages; or if you are buying for a team that needs SSO and a console.

Buy Coii VoiceInput if the audio must not leave your Mac; if it has to work with the network off; if you want your own words rather than an improved version of them; if a sentence lost to a stalled request is a real cost to you; and if $19 once suits you better than $96 or $288 a year.

The rest of the field, briefly: Wispr Flow is the other large hosted option, superwhisper and BetterDictation are the local Mac apps, MacWhisper is the one that also handles recordings, and Apple Dictation is free and already on your machine.

Everything this app connects to is a short list, and the trial is thirty days.