# logia
Logia is a small native GNOME voice-to-text app, written in Rust with GTK 4 and libadwaita. Press a global shortcut, speak, and the transcript lands in whatever app you were typing in — with a live waveform, persistent top-bar controls and a main window for models, settings, transcripts and corrections. All speech recognition runs on your computer; nothing is sent anywhere.
The engine is deliberately swappable, because no single model is best for every accent, language or device. The default is Parakeet v3 in INT8, which runs on the CPU with ONNX and automatically detects 25 European languages, English and Dutch included. Whisper via whisper.cpp is one click away, and Qwen3-ASR is an optional heavier alternative. Models are downloaded with pinned SHA-256 checksums and transcription always stays offline; only the explicit model download touches the network.
The feature I actually use daily: corrections. If Logia mishears a word, fix it where it was typed and Logia learns the phrase replacement locally — no model retraining, just phrase rules with case and punctuation handled properly. In GTK apps and LibreOffice this even works from the correction you make after pasting, via accessibility text events, with password fields strictly off limits. There’s push to talk, an optional history that is off by default, and recordings are deleted when a dictation finishes or is cancelled.
Install the native app with the installer script, or grab a Flatpak bundle built for every push. The app is MIT licensed; model weights keep their own licenses. If dictation in every app with a local model sounds like your thing, have a look at the forge.
~$