Overview

Takumi Voice

Hold a key, say what you mean, and clean text is typed where your cursor already was. These pages cover what it does, what it deliberately doesn't, and where every byte goes.

The loop

1. Hold your key. Recording starts. A small pill shows the level and the clock.
2. Speak. The pill never takes focus, so the app you're in keeps the cursor.
3. Release. The audio gets transcribed, then cleaned up for wherever it's going.
4. It's typed. We borrow your clipboard for about the length of a paste, then put it back.

What it isn't

No live transcript. Words appearing one at a time while you're still talking is something to watch, not something to use. The transcription request doesn't even ask for segment timings, since we'd only throw them away.

Nothing runs between sessions. No pill window exists, no timer is armed and no frame is drawn when you're not dictating. The hotkey watcher is an OS callback, not a poll. One exception: bare modifiers under X11, which X can't deliver any other way.

No provider account, model picker or API key. Transcription and cleanup both run server-side in a single request, so there's no second round trip between releasing the key and seeing words. Which models run is a service decision, changed without an app release.

Cleanup can't cost you a dictation. If it errors, times out, answers you instead of editing you, or invents code you didn't dictate, you get the raw transcript instead.

Where to go next

• Setup — permissions, choosing your key, first dictation.
• Dictation — formatting styles, editing by voice, languages.
• Vocabulary and snippets — teach it your names and phrases.
• Licence and machines — plans, seats, moving to a new laptop.
• What leaves your machine — the complete list.
• Troubleshooting — when the key does nothing.