A macOS desktop app that captures system audio, transcribes and translates it in real-time, and displays captions in a beautiful floating overlay.
100% local. 100% free. 100% private.
No API keys. No cloud services. No subscription costs. Everything runs on your machine.
- Real-time transcription and translation - See captions as people speak
- Multi-language support - Translate from any language to English (and more)
- Fast local processing - Uses optimized Whisper for speech-to-text
- Scrollable caption history - Scroll back to see previous translations
- Context-aware - Uses previous sentences for better accuracy
- Smart sentence combining - Handles "um", "ah" pauses gracefully
- Music detection - Shows indicator when music is playing (no speech)
- Beautiful overlay - Draggable, resizable, semi-transparent window
- Highly configurable - Adjust model size, chunk duration, appearance
- Higher beam search - Default beam_size increased from 1 to 3 for significantly better transcription accuracy
- Larger context window - Now uses 7 previous sentences (up from 3) for better contextual understanding
- High-quality audio resampling - Uses scipy polyphase resampling instead of linear interpolation
- Enhanced translation prompts - Few-shot examples improve Ollama translation quality
- RMS-based silence detection - More robust speech/silence classification
- Quality Profiles - Choose from Fast, Balanced, or Accurate presets in settings
- Streaming Translation - See translations appear word-by-word as they're generated
- Audio Normalization - Automatic gain adjustment for quiet audio sources
- Caption Export - Save your caption history to text or SRT subtitle files
- Keyboard Shortcuts:
Ctrl+Shift+TorSpace- Toggle translation on/offCtrl+Shift+C- Clear historyCtrl+Shift+S- Export captionsCtrl+,- Open settingsEscape- Stop translation
| Profile | Beam Size | Chunk Duration | Best For |
|---|---|---|---|
| Fast | 1 | 2s | Live conversations |
| Balanced | 3 | 3s | Movies/videos (default) |
| Accurate | 5 | 4s | Important recordings |
System Audio -> BlackHole -> Whisper -> Captions
(VLC, Zoom, etc.) (virtual audio) (transcribe + translate)
For English output: Whisper handles both transcription AND translation in a single pass (fastest!)
For other languages: Whisper transcribes, then Ollama translates
brew install blackhole-16chThen set up Multi-Output Device:
- Open Audio MIDI Setup (Spotlight -> "Audio MIDI Setup")
- Click + -> Create Multi-Output Device
- Check both:
- Your speakers/headphones
- BlackHole 16ch
- Set BlackHole 16ch as the Master Device (dropdown at top)
- Go to System Settings -> Sound -> Output -> Select Multi-Output Device
brew install ollama
ollama pull llama3.2
ollama serve # Run in background# Clone and enter directory
cd translator
# Create virtual environment
python3 -m venv venv
source venv/bin/activate
# Install dependencies
pip install -r requirements.txtpython main.py- Click the Start button to begin capturing and translating
- Play audio from any app (VLC, YouTube, Zoom, etc.)
- Watch captions appear in real-time
- Scroll up to see previous translations
- Drag the overlay to reposition
- Resize by dragging the corner
- Click the gear icon for settings
- Click the save icon to export captions
Edit config.json or use the Settings dialog:
{
"target_language": "English",
"whisper_model": "small",
"whisper_backend": "faster-whisper",
"quality_profile": "balanced",
"whisper_beam_size": 3,
"whisper_best_of": 2,
"audio": {
"sample_rate": 44100,
"chunk_duration": 3.0,
"silence_threshold": 100
}
}| Model | Size | Speed | Accuracy | Recommendation |
|---|---|---|---|---|
tiny |
75MB | Fastest | Low | Low-end hardware |
base |
145MB | Fast | Medium | Quick testing |
small |
488MB | Medium | High | Default - good balance |
medium |
1.5GB | Slow | Very High | High accuracy |
large-v3 |
3GB | Slowest | Best | Absolute best (needs GPU) |
large-v3-turbo |
809MB | Fast | Near-best | Best quality (transcription only) |
New: large-v3-turbo - Released by OpenAI in late 2024, this model offers near-best accuracy at 8x the speed of large-v3, with only half the memory requirements (~6GB VRAM vs ~10GB). It's the sweet spot for users who want high quality without the slowdown of large-v3.
Note: The turbo model only supports transcription, not Whisper's built-in translation. When using turbo with non-English audio, translation is handled by Ollama (requires Ollama to be running).
| Duration | Latency | Accuracy | Feel |
|---|---|---|---|
| 1.5s | Low | Lower | Fast but choppy |
| 2.0s | Medium | Good | Balanced |
| 3.0s | Higher | Better | Recommended |
| 4.0s+ | High | Best | Feels delayed |
faster-whisper- Recommended, works on all platforms, very stablemlx- Apple Silicon only, known to cause crashes - NOT RECOMMENDED
The app passes up to 7 recent transcriptions to Whisper as context, improving:
- Recognition of technical terms
- Consistency of names and proper nouns
- Overall coherence
Automatically combines sentence fragments:
- "like a drum?" + "Then we can tune it" -> Combined if no complete sentence
- Won't break on filler words like "um", "ah", "you know"
Automatically boosts quiet audio sources (up to 10x gain) for better transcription of low-volume content.
Detects and filters common Whisper hallucinations like "Thank you for watching", music notation symbols, and other artifacts.
Only inserts paragraph breaks when:
- Previous sentence ends with
.?! - AND there's 5+ seconds of silence
# Check BlackHole is installed
brew list blackhole-16ch
# Verify in Audio MIDI Setup that Multi-Output Device includes BlackHoleMake sure your app (VLC, etc.) is using the system audio output, not a specific device.
- Use a smaller Whisper model:
"whisper_model": "base" - Switch to "Fast" quality profile in settings
- For English output, Ollama is skipped automatically
- Reduce chunk duration to 2.0s (less accurate but faster)
This was fixed. If it happens, restart the app.
First run downloads the model (~500MB for small). This is a one-time download.
translator/
├── main.py # App entry point
├── config.json # User configuration
├── requirements.txt # Python dependencies
├── tests/ # Unit tests
│ └── test_improvements.py # Tests for v1.1 features
└── src/
├── audio/
│ └── capture.py # Audio capture via BlackHole
├── providers/
│ ├── base.py # Provider interface
│ └── translator.py # Whisper + Ollama integration
├── ui/
│ ├── overlay.py # Floating caption window
│ ├── settings.py # Settings dialog
│ └── tray.py # System tray (optional)
└── utils/
└── config.py # Configuration management
To create a standalone .app bundle you can double-click:
# Install build dependencies
pip3 install pyinstaller pillow
# Create the icon
python3 scripts/create_icon.py
# Build the app
pyinstaller Translator.spec --noconfirm
# The app will be in dist/Translator.app
open dist/Translator.appTo install to Applications:
cp -r dist/Translator.app /Applications/# Run all tests
python -m pytest tests/ -v
# Run only the v1.1 improvement tests
python -m pytest tests/test_improvements.py -v- macOS 12+ (tested on macOS 14+)
- Python 3.10+ (for development/building only)
- BlackHole 16ch virtual audio driver
- ~1GB disk space for app + Whisper model
- PyQt6 - GUI framework
- faster-whisper - Optimized Whisper implementation
- ollama - Local LLM client
- sounddevice - Audio capture
- numpy - Audio processing
- scipy - High-quality audio resampling
MIT License - Use freely!
- faster-whisper - Fast Whisper implementation
- Ollama - Local LLM runner
- BlackHole - Virtual audio driver
- PyQt6 - GUI framework
- Added Whisper large-v3-turbo model support (8x faster than large-v3, near-best accuracy)
- Note: turbo only supports transcription, not translation (uses Ollama for translation)
- Updated MLX backend warnings (known crash issues, not recommended)
- Documentation improvements for model selection
- Added quality profiles (Fast/Balanced/Accurate)
- Increased default beam_size from 1 to 3 for better accuracy
- Added streaming translation display
- Added caption export to TXT/SRT files
- Added keyboard shortcuts
- Improved audio resampling with scipy
- Added RMS-based silence detection
- Added audio normalization for quiet sources
- Enhanced Ollama prompts with few-shot examples
- Expanded context window from 3 to 7 sentences
- Added comprehensive unit tests
- Initial release