Whisper to your Omarchy.
Hold Right Alt. Say the thing. Let go. Your words land where your cursor was. That's it. That's the app.
asmir is push-to-talk dictation built for Omarchy and nothing else on purpose. It listens through a tiny waveform at the bottom of the screen, dressed in your theme, and turns speech into text right on your machine with whisper.cpp. No cloud, no account, no subscription. Your rambling stays yours.
Why "asmir"? It's ASMR for your keyboard: you whisper, it listens. And yes, it runs on whisper.cpp. We couldn't help ourselves.
git clone https://github.com/raffsmagic/asmir
cd asmir/pkgbuild && makepkg -si # builds the latest release into a real package
asmir setup # fetch the model and plug asmir into Omarchy's dictationThe AUR package is on its way: the AUR has paused new account sign-ups for
now. Once it's up, this becomes yay -S asmir.
asmir setup does the whole dance: downloads the speech model (32 MB, checked
against its SHA-256), adds asmir to Omarchy's Setup > Defaults > Dictation menu
next to the other backends and picks it there, has it start at login, and
reloads Hyprland. Then:
| Key | Does |
|---|---|
| Hold Right Alt | Talk while held, paste when you let go |
| Alt + Space | Start, then press again to stop and paste |
A recording you forget about stops and pastes on its own after five minutes.
To back out of one without pasting, bind a key to asmir cancel.
Changed your mind? Pick another backend under Setup > Defaults > Dictation. No hard feelings.
git clone --recursive https://github.com/raffsmagic/asmir
cd asmir
bin/dev-install # builds and installs into ~/.local/bin
asmir setup- Omarchy. Other Hyprland or layer-shell setups work too, minus the conveniences (see Off the beaten path).
- A processor from this decade, roughly: x86-64 with AVX2, meaning Intel from 2013 on (Haswell and newer) or AMD from 2015 on, every Ryzen included. That's what keeps whisper quick. ARM isn't supported yet.
- A microphone. asmir uses whichever one is your default right now, so plugging in a headset mid-session just works.
- Your theme. Accent, background, window rounding and font come from Omarchy, and the overlay follows a theme switch live.
- Your language. Messages come in all 50 of Omarchy's languages and follow Setup > Language as you change it.
- Your keys. asmir plugs into
omarchy-dictation, so Omarchy's Right Alt and Alt + Space work out of the box, and so does anything else that calls it. - Your paste. Text goes in through Omarchy's own paste, which knows a terminal wants Ctrl+Shift+V.
- Out of your way. The overlay is a layer-shell surface: it shows on the screen you're using and never steals focus or clicks.
- Honest about the mic. A microphone that's been idle needs a moment to wake up. Until the first real sound arrives, a faint glint runs across flat bars. When they come alive, talk.
The defaults are the point. If you insist, ~/.config/asmir/config takes:
model = tiny.en-q5_1 # base.en-q5_1 or small.en-q5_1: more accurate, slower
language = en # "auto" with a multilingual model like small-q5_1
prompt = Omarchy, Hyprland, Neovim. # words and spellings to expect
threads = 0 # 0 = half your CPU threads, up to 8
trailing_space = true # back-to-back dictations don't run togetherChanges apply from the next dictation, including a model you've just
downloaded with asmir model download.
| Command | What it does |
|---|---|
asmir setup |
Make asmir Omarchy's dictation backend |
asmir start / stop / toggle |
Record, then paste what was said |
asmir cancel |
Stop recording and paste nothing |
asmir status [--json] |
idle, recording, transcribing, or stopped |
asmir quit |
Stop the daemon |
asmir serve |
Run the daemon in the foreground |
asmir transcribe <file> [--json] |
Print what's said in an audio file (needs ffmpeg) |
asmir model |
The model in use and the ones on offer |
asmir model download [name] |
Fetch a model, checksum and all |
On another Wayland compositor with layer-shell (Hyprland, Sway, KDE Plasma,
niri), bind asmir start to a key press and asmir stop to its release.
Without Omarchy's paste command, asmir leaves the text on the clipboard and
tells you so. GNOME doesn't do layer-shell, so the overlay won't show there.
- C++ and Qt 6, built like Omarchy's own apps, with the overlay in QML.
- whisper.cpp 1.9.5, pinned as a submodule and compiled into the binary.
- A quantized Whisper model (
tiny.en-q5_1): a third of the full size, and fast enough that the text lands about as soon as you let go. - Up to 4× faster than other dictation on Omarchy, using less CPU and memory.
- A small daemon keeps the model loaded between dictations, and only borrows whisper's working memory while it transcribes.
- Recording straight through PipeWire: the connection stays open, and a capture stream exists only while you hold the key, on your default mic.
- GPU acceleration through Vulkan, for the bigger and more accurate models.
- ARM, for Omarchy on Apple Silicon and Snapdragon.
- A way for voice assistants and other tools to use asmir as their local ears.
Want one of these, or something else? See CONTRIBUTING.md.
bin/build # build/app/asmir
bin/test # unit tests, plus a real transcription when the model is downloaded
bin/bench # speed and accuracy on 73 LibriSpeech clips (needs uv)ASMIR_DRY_RUN=1logs the transcript instead of pasting it.ASMIR_DEBUG=1prints timings and whisper.cpp's own log.ASMIR_NATIVE=1 bin/build-whispertunes whisper.cpp for the building CPU only (deletebuild/whisperfirst).
MIT. See LICENSE. Standing on the shoulders of whisper.cpp and the Omarchy app templates.
