macOS · M1 and later
Apple Silicon
The one to take if About This Mac says M1, M2, M3 or M4. The engines run on the GPU through Metal.
Download .dmgCoii-VoiceInput_aarch64.dmg
Download
Pick your machine.
Every engine and every feature is unlocked for 30 days, with no card and no account. After that, starting a new dictation is the only thing that asks for a licence.
macOS · x86_64
Intel
For Macs made before Apple Silicon. Same app; the smaller engines are the sensible choice here.
Download .dmgCoii-VoiceInput_x86_64.dmg
Linux · x86_64
Debian / Ubuntu
Installs with apt and declares what it needs. Ubuntu 25.04 or later on Wayland; on 24.04, choose the Xorg session at login. The first launch asks once to allow the paste.
Download .debCoii-VoiceInput_x86_64.deb
Linux · x86_64
AppImage
One file, no package manager and no root. For everything that is not Debian or Ubuntu. The same desktop requirements as the .deb.
Download .AppImageCoii-VoiceInput_x86_64.AppImage
Not sure which? Apple menu → About This Mac. “Apple M1” or later means Apple Silicon; anything that says Intel means the other one. A browser cannot tell the two apart, which is why this page asks instead of guessing.
What the first run does
It asks for two permissions
Microphone, and Accessibility. The second is what lets the text be typed into another application rather than only copied. Each row turns green in front of you, so a refused permission is visible now instead of a mystery later. Linux has no Accessibility setting; under Wayland the desktop asks once instead, whether the app may send the paste.
It downloads one model
The recommended one, a few gigabytes, because it handles Chinese and English together. On a slow connection you can start with the light one instead and swap later. Downloads resume if interrupted and are checked against a hash.
It makes you dictate a sentence
Into a real text box, through the real injection path. The most common way a dictation app fails is silently, on the first try, in somebody else’s document. If it fails here, it fails with a diagnosis.
Requirements
- macOS — 13 Ventura or later, Apple Silicon or Intel.
- Linux — x86_64, glibc 2.35 or newer. On Wayland, a desktop with the global-shortcuts portal — GNOME 48 or later, which is Ubuntu 25.04 and newer — or any X11 session; Ubuntu 24.04 on Wayland has no dictation key, so choose its Xorg session at login. The text is pasted with Ctrl+V — on Wayland after a one-time permission from the desktop — so in a terminal, which pastes with Ctrl+Shift+V, take it from the history.
- Disk and RAM for the engine you choose, from a few hundred megabytes to a few gigabytes. One sentence is decoded at a time, so memory is the size of one model.
- Models, the transcript history and your vocabulary live under your own Application Support folder — ~/.local/share/CoiiVoice on Linux. Recordings are never kept. On macOS your trial and licence record is also mirrored in the login keychain, so deleting the folder does not reset the trial.
- The disk images are signed and notarised by Riff Tech, Inc., so the first open is an ordinary one. No right-click → Open.
$19 once when the 30 days are up.
Every engine, 3 computers, every new version at no further charge. Everything you dictated during the trial stays readable, searchable and copyable either way.