Beam Capture and type
Dictation
Hold a key, talk, let go. Cleaned-up text lands where your cursor is.
Triggers
| Trigger | How |
|---|---|
| Fn (default) | Hold Fn and talk. Set the globe key to Do Nothing in Keyboard Settings first. |
| Right ⌘ | Hold Right Command. |
| Hotkey | Press once to start and again to stop. Default is ⌥D when the trigger is set to hotkey. |
A compact pill with a live waveform and transcript appears near the bottom of the screen, with one-tap targets: cursor, Ask AI, or a note. Words you have already said are not lost if you pause.
Where text goes
Settings → Voice → Deliver to: Cursor (pastes), Ask AI, or Note.
Cleanup
On by default when AI is on: adds punctuation, removes filler words, no rewriting. Styles: None, Email, Messaging. Smart uses your main chat model, Fast uses the quick model. If a model refuses or answers instead of editing, Beam tries the other, then falls back to simple local rules.
Vocabulary
Add names and terms, comma separated, spelled the way you want. Beam corrects near-misses for terms of four or more letters.
High accuracy mode
Off by default. Needs an OpenAI or Groq key. Beam records audio and uploads it to OpenAI (gpt-4o-mini-transcribe) or Groq (whisper-large-v3) instead of using Apple speech recognition.
What stays on the Mac
Standard dictation uses Apple Speech and asks for on-device recognition when your language supports it. The code cannot tell you which languages fall back to Apple's servers. The cleanup step sends the transcript text to your configured AI model, which is local unless you chose a cloud one. High accuracy sends audio to the provider you picked.
Permissions: Microphone and Speech Recognition, plus Input Monitoring and Accessibility for hold triggers and pasting. Voice settings include diagnostics and test buttons.