voicecom

A macOS menu bar app for system-wide voice-to-text. Press a hotkey, speak, and the transcribed text is pasted into whatever app you're using. All processing happens on-device using Whisper or NVIDIA Parakeet models -- no data leaves your Mac.

Features

• Menu bar app -- lives in your menu bar, works system-wide with no dock icon
• Two input modes -- toggle recording (press to start, press to stop) or push-to-talk (hold to record, release to transcribe)
• Two model families -- OpenAI Whisper (via whisper.cpp) and NVIDIA Parakeet, selectable from a single model picker
• Hardware-accelerated transcription -- on-device inference with Metal GPU acceleration, CPU + Accelerate, and optional CoreML encoder for ANE (Whisper)
• Multiple model sizes -- Whisper from tiny to large-v3, plus Parakeet f16 and quantized variants, downloaded automatically on first use with a progress bar
• Language selection -- choose a transcription language or use auto-detect with multilingual models (Parakeet auto-detects the language)
• Automatic text insertion -- transcribed text is pasted directly into the frontmost app (full clipboard save/restore)
• Configurable hotkeys -- set your own keyboard shortcuts for toggle and push-to-talk
• Recording safety -- automatic stop after 5 minutes to prevent runaway recordings
• Fully offline -- models run locally, no network required after initial download

Screenshots

Menu bar popover	Settings — General	Settings — Permissions

Requirements

• macOS 26.2+
• Apple Silicon (ARM64) recommended for best performance
• Xcode 26+ to build from source

Getting Started

1. Clone the repository with submodules:

   git clone --recurse-submodules https://github.com/your-username/voicecom.git

2. Open voicecom.xcodeproj in Xcode

3. Build and run (Cmd+R)

4. Grant the requested permissions:

   • Microphone -- prompted on first recording
   • Accessibility -- needed to paste text into other apps (prompted at launch, or grant manually in System Settings > Privacy & Security > Accessibility)

Usage

Action	Default Shortcut
Toggle recording	Option + Shift + R
Push-to-talk (hold)	Option + Shift + T (enabled by default)

1. Click the mic icon in the menu bar or press the toggle hotkey to start recording
2. Speak
3. Press the hotkey again (or release for push-to-talk) to stop and transcribe
4. The transcribed text is automatically pasted at your cursor

Shortcuts can be changed in Settings (Cmd+,).

Transcription

voicecom supports two on-device model families, both powered by whisper.cpp (vendored at v1.9.1). Pick a model in Settings and the app swaps backends transparently -- Whisper and Parakeet models appear side by side in the same picker. Models are downloaded from HuggingFace on first use (with download progress shown in the UI) and cached under ~/Library/Application Support/voicecom/models/.

Whisper (ggml-* models)

GGML .bin weights (and an optional CoreML encoder) are cached in models/whispercpp/. Hardware acceleration is layered for maximum performance:

• Metal GPU -- accelerates the decoder via the ggml Metal backend (enabled with flash attention)
• CoreML encoder -- downloaded alongside the model for ANE acceleration (falls back to CPU automatically)
• Accelerate -- used for CPU-side BLAS operations
• Performance cores only -- inference threads are pinned to P-cores to avoid latency from E-cores

Parakeet (parakeet-* models)

NVIDIA's Parakeet TDT 0.6B v3 is available via whisper.cpp's parakeet_* C API (whisper.cpp v1.9.0+), and parakeet-tdt-0.6b-v3-q8_0 is the default model for new installs. Weights are downloaded as a single GGML .bin from the ggml-org/parakeet-GGUF HuggingFace repo and cached in models/parakeet/. Three variants are offered:

Model	Precision	Approx. download
parakeet-tdt-0.6b-v3	f16	~1.2 GB
parakeet-tdt-0.6b-v3-q8_0	q8_0	~650 MB
parakeet-tdt-0.6b-v3-q4_k	q4_k	smaller still

Parakeet runs on the Metal GPU (with CPU fallback) and, like Whisper, pins inference threads to P-cores. There is no CoreML/ANE path. Parakeet v3 is multilingual and auto-detects the spoken language, so the language setting is ignored while a Parakeet model is selected.

Language selection

You can select a transcription language in Settings (General tab). Use "Auto-detect" with multilingual Whisper models (e.g. ggml-small, ggml-large). English-only models (.en suffix) always transcribe in English, and Parakeet models always auto-detect, regardless of this setting.

Building from Source

# Debug build
xcodebuild -scheme voicecom -configuration Debug -derivedDataPath build/DerivedData

# Release build
xcodebuild -scheme voicecom -configuration Release -derivedDataPath build/DerivedData

# Run tests
xcodebuild test -scheme voicecom -derivedDataPath build/DerivedData

The Release .app bundle is output to build/DerivedData/Build/Products/Release/voicecom.app.

Creating a .dmg for Distribution

To package the app into a .dmg installer with a drag-to-Applications layout:

# 1. Build in Release mode
xcodebuild -scheme voicecom -configuration Release -derivedDataPath build/DerivedData

# 2. Create a staging directory
mkdir -p build/dmg_staging
cp -R build/DerivedData/Build/Products/Release/voicecom.app build/dmg_staging/
ln -s /Applications build/dmg_staging/Applications

# 3. Create the .dmg
hdiutil create -volname "voicecom" \
  -srcfolder build/dmg_staging \
  -ov -format UDZO \
  build/voicecom.dmg

# 4. Clean up staging directory
rm -rf build/dmg_staging

The resulting build/voicecom.dmg can be shared and opened on any compatible Mac. Users open the DMG and drag voicecom.app into the Applications folder to install.

Note: For distribution outside your own machines, sign with a Developer ID certificate and notarize:

codesign --force --deep --sign "Developer ID Application: Your Name (TEAM_ID)" build/DerivedData/Build/Products/Release/voicecom.app
xcrun notarytool submit build/voicecom.dmg --apple-id YOUR_APPLE_ID --team-id TEAM_ID --password APP_SPECIFIC_PASSWORD --wait
xcrun stapler staple build/voicecom.dmg

Project Structure

voicecom/
  voicecomApp.swift          # App entry point (MenuBarExtra + Settings scene)
  AppState.swift             # Central @Observable state, settings, service wiring
  Services/
    TranscriptionBackend.swift   # Protocol for transcription backends
    TranscriptionService.swift   # Facade that routes models to the right backend
    WhisperCppBackend.swift      # whisper.cpp (GGML) Whisper backend
    ParakeetBackend.swift        # NVIDIA Parakeet backend (Metal GPU)
    ModelDownloader.swift        # Model downloads with progress reporting
    AudioRecorder.swift          # AVAudioRecorder-based recording
    TextInsertionService.swift   # Clipboard + CGEvent paste
    HotkeyManager.swift         # Global keyboard shortcut handling
    PermissionManager.swift      # Mic + Accessibility permission checks
  Views/
    MenuBarView.swift            # Menu bar popover UI
    SettingsView.swift           # Settings window (General + Permissions tabs)
    HotkeyRecorderView.swift     # Keyboard shortcut capture widget
LocalWhisper/                # Local SPM package wrapping vendored whisper.cpp (Whisper + Parakeet)
  include/                   # Public C headers (whisper.h, parakeet.h) + module.modulemap
  MetalObjC/                 # Metal backend ObjC files (compiled without ARC)
  vendor/                    # whisper.cpp source (git submodule, v1.9.1)
  Package.swift              # Builds whisper.cpp for Apple platforms (CPU + Metal + CoreML)