MetaWhisp
Open-source voice-to-text for macOS. Press a key, speak, text appears at your cursor.
MetaWhisp is a free, open-source native macOS menu bar app. It transcribes speech on your Mac with OpenAI's Whisper through WhisperKit and pastes the text into the app you're typing in. With on-device transcription and the local AI model, nothing leaves your Mac. Cloud features are optional; Privacy lists what each one sends and where.
How It Works
- Press Right Command from any app
- Speak
- Press Right Command again
- Text is automatically pasted at your cursor
That's it. Works in any app that accepts paste: editors, browsers, terminals, messengers. Prefer holding the key while you talk? Switch to push-to-talk in Settings.
Features
Transcription
- Whisper large-v3 on your Mac's GPU and CPU, in WhisperKit's turbo build (about 3.2 GB, downloaded once). The Base model (about 150 MB) installs first, so you can dictate while the big one downloads.
- Tiny, Base and Small are there if you want a lighter model
- 100 languages with auto-detection (99 on the smaller models), or pin one of 11 in Settings
- Hallucination filtering (YouTube artifacts, silence patterns, text mixing three or more scripts); flagged text goes to the clipboard instead of being pasted
Text Processing
- Raw — verbatim transcription (the default; a Pro subscription switches it to Structured once)
- Clean — removes Russian filler words, offline
- Structured — AI-powered cleanup with grammar fixes and formatting
- The AI runs on an optional local model (Phi-4 Mini, or Apple Foundation Models on macOS 26 with Apple Intelligence on), on your own OpenAI or Cerebras key, or on Pro
Translation
- Right Option tap — record and translate
- Right Option hold (1.5s) — translate selected text in any app
- 11 languages: EN, RU, ES, FR, DE, ZH, JA, KO, PT, IT, UK, plus a Gen Z slang rewrite
- Uses the same AI as Structured mode
Meetings and Calls (off by default)
- Records your microphone and the Mac's sound on the Mac, labelled Me and Them
- With Auto-detect calls and Auto-start turned on, spots Zoom, Teams, Meet and other calls and starts recording by itself
- A recap after the call (decisions, action items, next steps) when an AI is set up
- With your own Deepgram or Gemini key, everyone on the call gets their own name in the transcript
MetaChat and Memory
- Ask about what you said: dictations, meetings, tasks and memories (needs an AI)
- In toggle mode (the default), hold Right Command to ask out loud; answers are read aloud (natural voices with Pro)
- Tasks from promises in your conversations, memories (off by default) and goals you set
- Pro adds a daily summary, weekly patterns (off by default) and search by meaning
Screen Context (off by default)
- Reads text from the window in front, in every app except the ones you exclude (or only the apps you list), so MetaChat can answer about it
- The Apple Passwords, Keychain Access, 1Password and Bitwarden apps are never read
Keyboard Layout Fix
- Fixes a word typed in the wrong layout (English ↔ Russian); on by default, needs the Accessibility permission
Smart Dictionary
- Learns a correction once you make the same edit twice
- Built-in brand recognition (44 brands)
- Text snippets (e.g., "my email" expands to your actual email)
- Fuzzy matching with case preservation
Dashboard & Analytics
- Words, transcriptions, translations, WPM, time saved
- Activity charts by day/week/month
- Records: streak, best day, longest recording
System Integration
- Lives in the menu bar; a Dock icon shows only while the main window is open
- Floating recording pill overlay (5 styles)
- Sound presets (Default, Bass, Signature) or your own sounds
- Toggle (the default) or Push-to-Talk mode
- An MCP server (metawhisp-mcp, built from source, not in the DMG; turn on MCP Server in Settings) lets Claude Desktop, Cursor and other MCP clients search your memories, tasks and conversations
- Obsidian export; indexing of Apple Notes, Calendar and folders you pick
- Auto-updates via Sparkle
Requirements
- macOS 14+ (Sonoma)
- Apple Silicon (M1 or later)
- About 60 MB for the app, plus about 3.2 GB for the default model (the smallest model is under 100 MB)
Building from Source
# Clone
git clone https://github.com/MetaWhisp/MetaWhisp.git
cd MetaWhisp
# Build (release mode for ML performance)
swift build -c release
# Create the signed app bundle (needs a Developer ID certificate)
bash build.sh
# The bundle is at .build/arm64-apple-macosx/release/MetaWhisp.app
# To install a build over /Applications/MetaWhisp.app, keeping the replaced one for --rollback: bash hot-swap.sh
Dependencies
Resolved automatically via Swift Package Manager:
以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗
