Hush 🤫

Hush is a privacy-first voice dictation tool for macOS. Inspired by Wispr Flow, it allows you to speak into any text field by holding the Fn key.

Everything happens locally on your device using the Apple Neural Engine. No audio is ever sent to the cloud.

Features

• ⚡️ Fast Transcription: Uses Whisper tiny.en model with 1-2s total latency for typical recordings
• 💎 Liquid Glass UI: macOS Tahoe-style translucent recording overlay using .glassEffect() API
• 🔒 Private & Local: All inference runs on-device using whisper.cpp and CoreML. Your voice data stays on your machine
• 🍎 Native: Built in Swift & SwiftUI with multi-threaded processing for Apple Silicon
• 📝 Universal: Works in any application—Notes, VS Code, Browser, Slack, Terminal, etc.
• 📋 Auto-paste: Transcribed text is automatically pasted at your cursor position

Requirements

• macOS: 26.0 (Tahoe) or higher
• Hardware: Apple Silicon (M1/M2/M3/M4) for CoreML acceleration
• Permissions:
  • Accessibility (to listen for the Fn key)
  • Microphone (to capture audio)

Installation

1. Clone the Repository

git clone https://github.com/Ang-Andrew/hush.git
cd hush

2. Download Models

Download the Whisper tiny.en model (optimized for speed):

mkdir -p ~/.hush/models
cd ~/.hush/models

# Download tiny.en model (~74MB)
curl -L -o ggml-tiny.en.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en.bin

# Download CoreML encoder for GPU acceleration (~14MB)
curl -L -o ggml-tiny.en-encoder.mlmodelc.zip https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en-encoder.mlmodelc.zip
unzip ggml-tiny.en-encoder.mlmodelc.zip
rm ggml-tiny.en-encoder.mlmodelc.zip

3. Build & Run

swift build
.build/arm64-apple-macosx/debug/Hush

Note: On first run, macOS will prompt you to grant Accessibility and Microphone permissions. You must enable Accessibility in System Settings for the app to detect the Fn key.

Usage

1. Launch Hush - You'll see a waveform icon in your menu bar
2. Place your cursor in any text input field
3. Press and HOLD the Fn key - A liquid glass overlay appears at the bottom of your screen
4. Speak your message
5. Release the Fn key - Text is transcribed and automatically pasted
6. View History - Click the menu bar icon to access previous transcriptions

Performance

Typical latency on Apple Silicon (M1):

• Short (1-4s): ~0.6-1.0s transcription, ~1.0-1.5s total
• Medium (5-7s): ~0.8-1.4s transcription, ~1.2-1.9s total
• Long (10-20s): ~1.8-2.7s transcription, ~2.2-3.2s total

Total latency includes transcription + paste time.

How It Works

Hush uses whisper.cpp, a high-performance C++ port of OpenAI's Whisper model.

1. Input Monitoring: CGEventTap detects Fn key press/release globally
2. Audio Capture: AVAudioEngine records 16kHz mono audio while key is held
3. Inference: Audio processed locally via CoreML (Apple Neural Engine) with multi-threaded optimization
4. Injection: Text automatically pasted at cursor via system clipboard

Technical Details

• Model: Whisper tiny.en (39M parameters, English-only)
• Audio Format: 16kHz, mono, Float32 PCM
• Threading: 2 threads optimized for tiny model
• UI: SwiftUI with native .glassEffect(.clear) material
• Concurrency: Swift 6 with strict concurrency checking

License

MIT License. Feel free to use, modify, and distribute.