Dictate prompts to AI coding agents with local voice-to-text

Think about the last time you explained a bug to a colleague. It took a minute, out loud, and they got it. Now think about the prompt you typed to your AI agent for the same bug. It was probably three lines.

That gap matters. A good prompt for a coding agent is long. It names the files, describes the bug, lists what you already tried and says what "done" looks like. Typing all of that is slow, so prompts get short, and short prompts make agents guess.

Voice-to-text closes the gap. You talk the way you would to a colleague, and the agent gets the full story.

In short

• Hold a key, speak, release. The text lands where your cursor is, including a terminal.
• Transcription runs on your computer. No audio leaves it.
• Pick Parakeet on Apple Silicon or small.en with whisper.cpp for English.
• Recordings can run up to 10 minutes, and double-tap gives you hands-free mode.

What developers need from dictation

Generic dictation tools often fall short for coding work. Three things matter most:

How Voice-to-Text works in DevWorkspace Pro

Voice-to-Text in DevWorkspace Pro dictates text into any app. That includes your terminal with Claude Code, your editor, your browser and your chat.

Everything stays on your computer

Transcription runs locally. No audio, transcripts or metadata go to external servers. The only network access is the one-time model download.

The speech engine ships inside the app on every platform. There is nothing to compile or install.

Hold, speak, release

1. Hold the activation key. The default is right ⌘ on macOS and right Ctrl on Windows and Linux.
2. Speak. An overlay shows at the bottom of the screen.
3. Release. The text goes where your cursor is.

Recording starts only after you hold the key for the Hold Duration (default 300 ms). A quick tap does not record. If you press another key or click while you hold the activation key, the app treats it as a shortcut (for example right ⌘ + Enter) and does not record.

Got a long prompt? Double-tap the key for hands-free recording, and tap it again to stop. You can type and click while it records.

Built to paste into terminals

Insert via clipboard is on by default. The app puts the text on the clipboard and pastes it with ⌘V or Ctrl+V, which works well in terminals and code editors. Turn it off and the app types each character one at a time instead.

Keep transcript on clipboard decides whether the app restores your previous clipboard after the paste.

If the app cannot type into the focused app, it copies the text to the clipboard. The overlay then shows Copied and the reason.

Pick a model

Apple Silicon (Neural Engine):

whisper.cpp (all platforms):

An English-only (.en) model is more accurate than a multilingual model of the same size. For other languages, use a multilingual model and set Recognition Language.

A simple workflow for spoken prompts

1. Open the AI launcher with ⌘J (Ctrl+J) in a project, or focus the terminal where your agent runs.
2. Double-tap the activation key for hands-free mode.
3. Talk through the task as you would to a colleague: what is wrong, where, what you tried, what done looks like.
4. Tap the key to stop. Read the text and fix file names if needed.
5. Press Enter.

A few habits make the result much better:

• Say file and function names slowly. Then check them in the text. Code identifiers are the hardest part for any speech model.
• Paste exact strings. For an error message or a stack trace, paste it and dictate the explanation around it.
• Lead with the goal. The first line of a prompt in the AI launcher becomes the session title, so say the goal first.

When the text lands in the wrong window

It happens. During a long transcription, you switch apps by mistake and the text goes somewhere else.

Hold the activation key to show the overlay, then click the copy icon. The app puts your last transcript on the clipboard. It keeps that transcript until you quit.

Limits to know

• One recording can be up to 10 minutes. The timer changes color 30 seconds before the limit.
• If the microphone disconnects, the app transcribes the audio that it has.
• The computer does not sleep during a recording.
• On macOS, the app needs the Accessibility permission to paste into other apps.
• Voice-to-Text is fully tested only on macOS. On Linux, it uses xdotool on X11 and ydotool on Wayland. On Wayland, the activation key works only when an X11 or XWayland window has focus.

Set it up in three steps

1. Open Preferences → Voice-to-Text and turn on Enable Voice-to-Text.
2. Select a Transcription Approach and a model.
3. Click Download on the model card.

You download each model once, and then you are ready to talk.