Hold the right ⌥, say what you mean, let go. The words appear in whatever app you were already in — the reply, the brief, the commit message, the field you were halfway through filling in. Hold the right ⌘ instead and the same sentence comes out translated.
In development · not yet released
Dictation tools usually ask you to leave what you are doing, speak into their window, then copy the result back. This one is a key you hold.
Hold it and talk. Let go and the text is inserted at the cursor, in the app that was in front. Nothing to open, nothing to paste, nothing to switch back from.
The same hold-and-talk, with a translation step on the way out. Speak Polish, an English sentence arrives. Translation is a separate key rather than something the app guesses at, which is the point — see below for why.
Started a sentence you did not want? Press Escape while still holding, and nothing is transcribed and nothing is inserted.
Insertion goes through a paste, so the clipboard is saved first and restored afterwards. Whatever you had copied is still there when the sentence lands.
Some apps will not take a synthetic paste. Those get the text typed in character by character instead, so the answer to "it does not work in this one app" is not a shrug.
A microphone needs a few hundred milliseconds before it delivers usable audio, and by then you are already talking. The microphone is kept warm between recordings and half a second of audio is held back, so the beginning of the sentence is there.
Every item here exists because it is a common, specific complaint about dictation tools — not because it looked good on a feature list.
The most common failure in this category is the hotkey dying mid-session and never coming back until a restart. The key is watched at a level that can be re-armed, and it is re-armed when the system switches it off.
On a short sentence, automatic language detection often picks wrong — and a wrong guess does not produce a typo, it silently produces a translation. So the dictation language is something you choose, and translating is the other key.
Speech models invent sentences when handed silence — subtitle credits, thank-yous, whole polite closings that you never said. Audio with nothing in it is not passed on, so it has nothing to invent from.
Loading a speech model onto Apple's Neural Engine can hang with no error and no timeout, at nought per cent processor, indefinitely. Loading has a deadline; when it passes, the app loads a different way and remembers to do that next time. An app that goes quiet for an unbounded time is indistinguishable from a broken one.
No dock icon to alt-tab past, no window in your way. A small indicator while it is listening, and otherwise nothing.
Distributed directly, with a Developer ID signature and Apple's notarisation, so it opens without the warning that a downloaded app usually shows.
Speech is turned into text by a model that runs on your own Mac, using the machine's own hardware. Your voice is not uploaded to us, and it is not sent to any speech service.
What the app does contact us about, once you have bought it, is your licence key — see the Privacy Policy.
The app is in development. Rather than describe it as though it were done, here is where it actually stands.
The speech model runs on the Neural Engine in Apple's own chips, so an Intel Mac is not supported.
Granted once, in System Settings, the first time you dictate.
Also granted once. They are what let the app notice the key you are holding and put text into another application — the two things it is for.
Under a gigabyte on disk, fetched once during setup.
No subscription, and the version you buy keeps working. Until then the most useful thing you can do is tell us what would make it worth switching to — that still changes what gets built.