Grant permissions
The wizard opens with a permissions screen, which it does not count in its own step numbering. The screen skips itself if you have already granted them, and it does not appear on Linux.
- Microphone (macOS and Windows). So SpeakoFlow can hear you.
- Accessibility (macOS only). So it can type into your other apps.
Pick a voice model
This is the model that turns your voice into text, and it runs entirely on your machine. A fast English model or a real-time multilingual model are both good starting points.The download runs in the background, so you don’t wait here. You can switch models anytime in Settings → Dictation, and the full catalog is in Transcription models.
Give it a brain (optional)
Download a local model for the assistant and AI cleanup, or choose Skip for now. Three Gemma 4 sizes are offered:
- Quick (E2B). Responds fastest, less capable on complex requests.
- Balanced (E4B, recommended). The best conversational quality-to-latency tradeoff.
- Capable (12B). More capable, but noticeably slower.
You're ready
The last screen shows your two shortcuts as keycaps, plus a status line for anything still downloading. Your hotkeys are already live.
Say your first sentence
Open any text field, hold your Dictate shortcut, and talk. Let go, and your words type themselves.
Every shortcut is rebindable in Settings → General → Shortcuts.
How the size badges are chosen, and what happens if a download fails
How the size badges are chosen, and what happens if a download fails
The Quick, Balanced, and Capable labels are editorial. They are not inferred from your hardware.Tapping a card starts the download and flips the footer straight to Continue, so you can move on while it finishes. The built-in assistant provider is only pointed at the model once the weights are actually on disk, so a failed or cancelled download can’t leave the assistant set to a model you don’t have.