Voice typing that never leaves your computer

No. The audio is recognised by a model file on your own disk and is deleted with the temporary folder it lived in. The only thing that ever crosses the network is the one-time model download; after that dictation works with the network off entirely.

780 milliseconds from releasing the key to the text appearing, measured on a real five-second phrase with the default model on an M5 laptop. A fifteen-second phrase took 0.66–0.75 seconds. The heavier large-v3-turbo model takes roughly twice as long.

Yes — the hotkey is registered system-wide, so it fires in a terminal, a browser, a chat client or a text editor, with MediaChef itself in the background or its window closed.

The right ⌥ key by default — a lone modifier: hold it and speak, or tap once to start and again to stop; with Shift the text is sent with Enter. A modifier on its own types nothing and clashes with nothing. Combos like ⌥ Space typed a space in hold mode when ⌥ was released first, so they were dropped; ⌃⌥ Space and ⌃⌥ D remain as fallbacks.

For two things. Hearing the trigger key: a lone modifier cannot be registered as an ordinary hotkey, so MediaChef listens to the keyboard itself, and macOS allows that only with Accessibility. And typing text into another application's window, which the system treats as synthetic input. One permission covers both and is granted once; restart the app afterwards.

The text goes to your clipboard instead and a notification says why, with the right pane of System Settings opened for you. Nothing dictated is ever lost to a missing permission.

The app plus one speech model: 488 MB for the default, 78 MB if you pick the smallest, 1.62 GB for the largest. If you already use MediaChef to transcribe files, the model is on your disk already and dictation adds nothing.

Whisper supports 99 languages, and you can either name yours or let it detect. Mixing languages in one sentence is the case where the heavier model earns its size, and where the dictionary helps most.

Five minutes, after which it stops on its own and transcribes what it heard rather than discarding it. In practice people dictate in sentences, not monologues.

Escape while recording throws the take away and delivers nothing. It is registered only while you are dictating, so it does not interfere with Escape anywhere else.

It does the same job with two differences that matter: it can be taught your vocabulary, and the audio stays on your machine. If neither matters to you, the built-in one is already there and free.

Yes. MediaChef is open source under GPL-3.0 with no paid tier and no subscription. The current release is 0.8.5, dictation included.

Those are two separate passes. While you speak, the panel shows a draft produced by a light model every second and a half, so it can keep up with speech. What gets typed is the final result from your main model over the whole recording, with your term dictionary applied. Both models are picked on the Dictation tab: set the live preview to the same model as the main one and the difference disappears — at the cost of the panel lagging behind.

macOS ties a permission to the app's code signature, not to its name. MediaChef has no Apple certificate — we are not paying for one — so every version is signed differently and the system treats the updated app as a new one. The app removes the stale entry itself and asks again: one toggle and one “Allow” per update.

Then the entry in that list belongs to the previous copy of the app, which is what happens after an update. Switching it off and on again does not rebind it — we tried. MediaChef clears that entry itself and triggers the system prompt: switch it on in the list and restart the app.

Usually it is the headset: AirPods sitting in their case stay connected and stay the default input, and macOS hands out flat zeros from them — to every app, not just ours. The notification names the device the silence came from. Put the headset on, or pick the microphone explicitly on the Dictation tab.

Yes — the Dictation tab lists the input devices. “System default” means the last connected headset rather than the laptop microphone, which is exactly why an explicit choice is more reliable.

Nothing is typed. An empty recording never even reaches recognition: Whisper does not stay quiet on silence, it invents — the Russian model would type out subtitle credits. Those subtitler signatures are recognised and treated as silence, and you get “no speech” instead.

The macOS audio subsystem is waking up — up to two seconds for the first microphone open per launch. The panel appears immediately, before recording starts, so hold the key until it says it is listening. After that, opening takes tens of milliseconds.

Press the trigger together with Shift: the text is typed and Enter follows automatically. An empty dictation never presses Enter — otherwise staying silent would send an empty message, or run the previous command in a terminal.

The recording is cancelled and nothing is typed. The right ⌥ plus a letter is somebody else's keyboard shortcut, not dictation, and the app treats it as one.

No. The recording lives in a temporary folder and is deleted with it, and keeping a history of transcripts is off by default. The app promises your content never leaves the machine; writing everything you dictate onto it in plain text would sit oddly with that — people dictate passwords and pieces of private conversations.

Yes. The panel is drawn above full-screen windows and on every desktop — full-screen editors and terminals are exactly where people dictate.

The Dictation tab also offers the right ⌘ — the other lone modifier macOS assigns no action of its own to — plus two ordinary combinations, ⌃⌥ Space and ⌃⌥ D, in case both modifiers are already taken.