OwlWhisper

Local voice input for macOS — hold a hotkey, speak, release, text appears at your cursor.
Fully offline. No cloud APIs. Chinese-optimized with FireRedASR2.

English | 中文

---

How It Works

1. Hold your hotkey (default: Fn)
2. Speak naturally
3. Release — text is transcribed and pasted at your cursor

Everything runs locally on your Mac. No data leaves your device.

Features

• Hold-to-speak — Global hotkey works in any app, fully configurable
• Offline ASR — FireRedASR2 int8 via sherpa-onnx C API (2.89% CER on Chinese)
• Auto punctuation — ct-transformer restores commas, periods, question marks
• Auto paste — Transcribed text is pasted at cursor via simulated Cmd+V
• VAD — Silero VAD trims silence, only speech is sent to ASR
• Floating indicator — Waveform animation while recording, dots while transcribing
• Auto model download — First launch downloads ~1.1GB of models automatically with resume support
• i18n — English (default) + Chinese
• Update checker — Checks GitHub Releases for new versions

Requirements

• macOS 13.0+
• Apple Silicon (arm64)

Install

From Release (recommended)

1. Download OwlWhisper-0.1.0.dmg from Releases
2. Open DMG, drag OwlWhisper to /Applications
3. Launch — models download automatically on first run (~1.1GB)

From Source

git clone https://github.com/sanvibyfish/OwlWhisper.git
cd OwlWhisper
scripts/setup.sh          # downloads native libs + models
open OwlWhisper/OwlWhisper.xcodeproj
# Build & Run (⌘R)

Permissions

On first launch, OwlWhisper will guide you through granting the required permissions:

Permission	Purpose
Microphone	Record your speech
Accessibility	Simulate Cmd+V paste and listen for global hotkeys

Architecture

OwlWhisper.app
├── ASRService.swift          # sherpa-onnx C API: VAD + ASR + punctuation
├── AudioRecorder.swift       # AVAudioEngine 16kHz mono capture
├── HotkeyManager.swift       # CGEventTap global hotkey detection
├── FloatingIndicator.swift   # Waveform + dots animation overlay
├── SettingsWindowController  # Auto Layout settings & onboarding
├── ModelDownloader.swift     # Chunked parallel download with resume
├── UpdateChecker.swift       # GitHub Release version check
├── MenubarController.swift   # Status bar icon + menu
├── PasteController.swift     # CGEvent Cmd+V simulation
├── AppDelegate.swift         # App lifecycle & orchestration
└── Frameworks/
    ├── libsherpa-onnx-c-api.dylib
    └── libonnxruntime.1.23.2.dylib

Models

Downloaded to ~/Library/Application Support/OwlWhisper/models/ on first launch:

Model	Size	Purpose
FireRedASR2 int8	~300MB	Chinese/English speech recognition
ct-transformer	~270MB	Punctuation restoration
Silero VAD	~2MB	Voice activity detection

Tech Stack

• Swift 5.9, AppKit (no SwiftUI)
• sherpa-onnx C API via bridging header
• onnxruntime 1.23.2

License

GPL-3.0