No visible UI
No window stack, no settings panel, no interruption. Koe lives as a tiny menu bar utility.
Koe
こえHotkey in. Speak naturally. Let text arrive polished.
Koe is a native macOS background voice input tool. Hold or tap a trigger key, stream speech to Doubao, let an OpenAI-compatible model clean the transcript, and paste the result straight into the focused app.
Plain-text control
Config, prompts, and dictionary all live in ~/.koe/ so everything stays scriptable and editable.
Zero Electron
Built with Objective-C for native macOS APIs and Rust for the core pipeline, without dragging an Electron shell into the app.
Small footprint
Under 15 MB on disk and typically around 20 MB memory, so it stays lightweight in the background.
One small utility. No giant app shell.
Koe currently targets Apple Silicon Macs and keeps the setup simple: install, grant three permissions, edit plain files, start dictating.
Click to copy
こえ, the Japanese word for voice.
Koe is written as こえ in hiragana and 声 in kanji. In Japanese it means voice, but not in a cold technical sense. It carries the feeling of something immediate and human: a thought leaving the body, crossing a tiny distance, and becoming language others can receive.
That is the product ideal behind Koe. It should feel light, direct, and almost invisible, more like a quiet bridge between speaking and writing than a separate app asking for attention.
Why it feels different
No GUI-first onboarding. You configure with files, not modal dialogs.
Dictionary entries are plain text, so terminology can be generated or edited in bulk.
Config changes apply on the next hotkey press instead of waiting for an app restart.
A four-step voice pipeline tuned for actual writing
Koe keeps provider support narrow on purpose so the full path from key press to pasted text stays understandable and fast.
Trigger instantly
Use Fn by default, or switch to another modifier key without touching a preference window.
Stream as you speak
Audio goes to Doubao ASR in real time so Koe can stay responsive instead of waiting for a full upload.
Correct the text
An OpenAI-compatible model fixes punctuation, spacing, terminology, and filler words using your prompts and dictionary.
Paste anywhere
The cleaned result is pushed back into the focused app with clipboard plus simulated paste automation.