Everything below happens on your computer. There is no Wrenscribe server in the loop.
1. The shortcut
GNOME tells Wrenscribe when your shortcut is pressed and released through the GlobalShortcuts portal. A hold records until you let go. A short tap starts hands-free recording until the next tap.
2. Recording
Audio is captured at 16 kHz, mono. A start sound plays and the pill appears with live levels. If you get near the recording limit (five minutes unless you change it) a warning sounds thirty seconds before the end.
3. Speech to text, while you talk
Wrenscribe does not wait for you to finish. At each pause it hands the stretch of speech so far to Parakeet TDT 0.6B, an open English speech model running on your CPU. By the time you let go, most of the transcript already exists.
4. Your words first
Before the language model sees anything, your dictionary replaces known mishearings and sound-alikes, then the built-in vocabulary fixes product and tech names the speech model is known to get wrong. Spoken symbols and commands are handled here too.
5. Cleanup
A small language model (Gemma 4 E2B, through llama.cpp) edits the transcript under a fixed set of rules: remove fillers and accidental repeats, apply self-corrections, keep your words and meaning, never change names, numbers or commands, turn counted items into lists, and follow the style for the app you are in. A fact guard then checks the result. If the model dropped a name or a number, added one, or wrapped the text in a preamble, Wrenscribe pastes the pre-model text instead. Built-in names are put back if the model changed them.
6. Save, then paste
The raw and the final text are written to history before anything is pasted, so a failed paste never loses a dictation. Then the GNOME Shell extension pastes into the focused app (Ctrl+Shift+V in terminals, Ctrl+V elsewhere) and restores your clipboard. If that cannot happen, the text is left on the clipboard and the pill tells you to press Ctrl+V.