Dictate with MacWhisper

Press a key, speak, press again. The words land at your cursor.

Local dictation for Visual Studio Code on macOS, transcribed by the MacWhisper app you already own. Project page: bitman.ch/en/dictate-with-macwhisper.

By default the extension makes no network requests: it records with ffmpeg and hands the audio to MacWhisper's command-line tool mw, which uses whatever engine you selected in the app. With a local engine (Parakeet, WhisperKit, Apple on-device speech) nothing leaves your Mac; if you selected a cloud engine in MacWhisper, MacWhisper sends the audio there, as it does for any transcription. The optional clean-up pass (off by default) sends the transcript text, never the audio, to the Ollama server you configure; with Ollama on your Mac it stays local, with a remote server the text leaves your Mac.

• One shortcut: ctrl+alt+space (⌃⌥Space) toggles recording.
• Inserts through the editor API: replaces the selection, works with multiple cursors. The clipboard is used only as a fallback when no editor can take the text, and a warning tells you.
• Works in Remote-SSH, WSL and dev-container windows (the extension runs on your Mac, the text goes into the remote editor).
• No API keys, no telemetry, no model downloads. A few hundred lines of TypeScript, shipped unminified in the extension package so you can read what runs.

Requirements

1. macOS with MacWhisper from macwhisper.com (the Mac App Store build has no mw tool).
2. The mw command-line tool: in MacWhisper open Settings → Advanced → Command-Line Tool → Install. The extension also finds the copy inside MacWhisper.app if you skip this step.
3. ffmpeg for microphone capture: brew install ffmpeg. macOS asks once to let Visual Studio Code use the microphone. If you declined, enable it under System Settings → Privacy & Security → Microphone; until then the extension reports "No audio was captured".
4. The shortcut must reach VS Code. If you use several keyboard layouts, macOS binds ⌃⌥Space to "Select previous input source" and swallows it: turn that off under System Settings → Keyboard → Keyboard Shortcuts → Input Sources, or rebind the command in VS Code (Keyboard Shortcuts → "Dictate with MacWhisper: Toggle").

Usage

While recording the status bar shows Recording…; while MacWhisper works it shows Transcribing…. A tap shorter than 0.3 s inserts nothing and the status bar says so for two seconds. A recording stops by itself after one hour, so a crashed window cannot leave the microphone open; the status bar and the macOS microphone indicator show that a recording is running. On an M-series Mac with Parakeet v3, ten seconds of speech come back in about a second.

If no text editor is active when the transcript arrives, or the editor is read-only, the text is copied to the system clipboard and a warning says so (clipboard managers and Apple's Universal Clipboard see it there). Text is never dropped silently. Each recording lives in a private temporary folder that is deleted when the dictation ends, whatever the outcome.

Optional clean-up: punctuation by a local model

Dictation comes out as one unpunctuated run. If you have Ollama on your Mac, the extension can send the transcript to a small local model that is asked to add punctuation and capitalization and nothing else:

brew install ollama && brew services start ollama
ollama pull gemma3:4b

Then set macwhisperDictate.cleanup.endpoint to http://127.0.0.1:11434. On an M4 Pro the pass took one to three seconds with a warm model. The answer is accepted only if it contains exactly the dictated words in the same order (case and punctuation aside) and no invisible, control or direction-changing characters; if the model changed, dropped, added or reordered a word, did not finish, answered with more than 1 MB, or did not answer within 20 seconds, the raw transcript is used and a warning says so. This check catches changed words, not a meaning changed by punctuation alone. The transcript goes only to that endpoint, over plain HTTP unless you give an https:// URL. Models with a thinking mode (qwen3) are slow here; gemma3:4b has none.

Settings

Why not just MacWhisper's own dictation hotkey?

Use it, it is great. MacWhisper's hotkey types into any app by simulating keystrokes. This extension exists for the cases where that is the wrong tool inside an editor: replacing a selection, multi-cursor, Vim mode and remote windows. It is also a toggle rather than hold-to-talk, which some people prefer for longer notes.

Questions and problems

Write to kontakt@bitman.ch or use the Q&A tab on the Marketplace listing. The project page with FAQ: https://bitman.ch/en/dictate-with-macwhisper/#support

Development

npm install
npm test            # core tests (uses the real mw and a Polish WAV fixture) + extension-host test
npm run package     # builds the .vsix

Core logic lives in src/core/ and has no dependency on the vscode module, so it is tested with plain node --test.

Not affiliated

MacWhisper is a product and trademark of Good Snooze. This extension is an independent, MIT-licensed companion and is not endorsed by them.

License

MIT