Voice-to-text transcription that pastes automatically at your cursor.

Starling is a lightweight macOS menu bar app that lets you dictate text anywhere with a global hotkey. Speak naturally, and when you stop, the text appears instantly at your cursor—no context switching required.

• Global Hotkey — Press ⌃⌥⌘J (customizable) to start/stop recording
• Voice Activity Detection — Automatically stops when you finish speaking
• Local Transcription — Uses Parakeet v3 Core ML (via FluidAudio) running on your Mac's Neural Engine
• Smart Paste — Automatically pastes transcribed text at your cursor without stealing focus
• Privacy First — No audio leaves your Mac; models cache locally in ~/Library/Caches/

brew tap Ryandonofrio3/starling
brew install starling

1. Download the latest .app from Releases
2. Drag Starling.app to /Applications
3. Open the app — it runs in your menu bar (look for the bird 🦜)

After download, it will warn you that the app is not from the App Store. Go to System Settings -> Privacy & Security and scroll down to Starling and click "Allow".

The app will guide you through a quick onboarding:

1. Microphone Access — Grant permission to record audio
2. Accessibility Access — Required to simulate paste (⌘V) and detect cursor position
3. Model Download — First run downloads the ~2.5 GB Parakeet v3 Core ML model (one-time, requires internet)

1. Press ⌃⌥⌘J (or your custom hotkey) to start recording
2. Speak — the bird glows while listening
3. Stop naturally — VAD detects when you finish, or press the hotkey again to stop manually
4. Text pastes automatically at your cursor (or copies to clipboard)

Access preferences from the menu bar bird icon:

• Hotkey — Customize the global shortcut (default: ⌃⌥⌘J)
• Trailing Silence Duration — Adjust VAD sensitivity (how long to wait after you stop speaking)
• Clipboard Retention — Keep transcription on clipboard after auto-paste (off by default)

• Local Processing — All transcription happens on your Mac via Core ML + Neural Engine
• No Network Calls — After initial model download, app works fully offline
• No Audio Storage — Audio buffers are processed in memory and immediately discarded
• Secure Input Respect — Password fields trigger copy-only mode (no keystrokes simulated)
• Model Cache — FluidAudio stores models in ~/Library/Caches/FluidAudio/ (~2.5 GB)

• macOS 14.0 (Sonoma) or later
• Apple Silicon (M1/M2/M3) or Intel Mac with Neural Engine support
• ~2.5 GB free disk space for model cache
• Permissions — Microphone + Accessibility access

• Xcode 15.0+
• macOS 14.0+ SDK
• Swift 6.0 toolchain

# Clone the repository
git clone https://github.com/Ryandonofrio3/starling.git
cd starling/Starling

# Open in Xcode
open Starling.xcodeproj

# Select the Starling scheme and build (⌘B)
# Run with ⌘R or archive for distribution

Or build from the command line:

cd Starling
xcodebuild -scheme Starling -configuration Release build

• Check Accessibility — System Settings → Privacy & Security → Accessibility → enable Starling
• Restart Required — After granting Accessibility, restart the app

• First launch only — Model download (~2.5 GB) requires a stable internet connection
• Check progress — Bird shows download status
• Clear cache — If download fails, quit app and run: rm -rf ~/Library/Caches/FluidAudio/

• Conflict detection — Another app may be using the same hotkey
• Change hotkey — Open Preferences and set a different combination
• Check permissions — Ensure Accessibility is granted

• Microphone check — Test your mic in System Settings → Sound
• Reduce background noise — VAD is sensitive to ambient sound
• Adjust trailing silence — Increase duration in Preferences if getting cut off

• English-first — Parakeet v3 supports 25 languages but is optimized for English
• No streaming — Transcription happens after you stop speaking (no live text yet)
• Large model — 2.5 GB cache requirement (no smaller English-only variant available)
• Offline-only — No cloud sync or multi-device support

• Streaming transcription (pending FluidAudio partial results support)
• Custom model management (clear cache, view size, switch versions)
• Press-and-hold mode (alternative to toggle)
• Launch at login option
• Multi-language selection UI
• Developer diagnostics panel (ANE usage, cache stats)

Contributions are welcome! Please open an issue or PR for:

• Bug fixes
• Feature requests
• Documentation improvements
• Performance optimizations

See CONTRIBUTING.md for guidelines.

MIT License — see LICENSE for details.

• FluidAudio — Swift wrapper for Parakeet ASR models
• NVIDIA Parakeet — Underlying speech recognition model
• Cursor Rules Contributors — Project structure and workflow inspired by community best practices

Have feedback or questions? Open an issue on GitHub.