Local voice-to-text dictation for Windows. Press a hotkey, speak, text appears. Runs fully offline on your NVIDIA GPU with OpenAI's Whisper large-v3, delivering near-perfect accuracy for English, German, and 90+ other languages.
Download ZIP | Requires Windows + NVIDIA GPU + Python 3.12+
- Hotkey dictation: Press
CTRL+ALT+D, speak, press again, text gets pasted into the active window - Offline & private: Everything runs locally on your GPU, no audio ever leaves your machine
- Fast: 73 seconds of speech transcribed in 7.7 seconds (9.5x real-time on RTX 4060)
- Accurate: CUDA float16 with beam search, handles dialects, background music, and long pauses
- Multi-language: Works with English, German, and all other Whisper-supported languages
- Dashboard: Click the tray icon to see today's stats, recent transcription history with click-to-copy, and quick actions (Calm Mode, restart, quit)
- Electric Border recording overlay: Animated microphone icon with dual-ring plasma effect (2D pixel displacement, breathing pulse, core flash), pre-rendered at 30fps. Red pulsing bar across all monitors
- Calm Mode: Toggle in the dashboard to replace the animated overlay with a simple static icon. Setting persists across restarts
- Spoken punctuation: Say "colon", "question mark" etc. and get the actual character (configurable)
- Hallucination filter: Known Whisper phantom outputs are detected and discarded
- System tray: Runs quietly in the background with a color-coded status icon (gray/green/red)
- Audio feedback: Beep tones on start/stop so you know when recording begins and ends
- History log: All transcriptions are saved with timestamps to
whisper-history.log - Autostart: Launches automatically on Windows login
- Single file: The entire tool is one Python script, easy to understand and customize
- Windows 10/11
- Python 3.12+
- NVIDIA GPU with CUDA support (tested on RTX 4060)
- Up-to-date NVIDIA driver
git clone https://github.com/TryoTrix/whisper-type.git
cd whisper-type
install.bat
The installer will:
- Check for Python, pip, and NVIDIA GPU
- Install all Python packages
- Create an autostart shortcut
- Download the Whisper model (~3 GB, one-time)
- Start the dictation tool
| Action | Shortcut |
|---|---|
| Start/stop recording | CTRL+ALT+D |
| Restart (if hook is lost) | CTRL+ALT+W |
Tray icon colors:
| Color | Status |
|---|---|
| Gray | Model loading (~4s) |
| Green | Ready |
| Red | Recording |
Left-click the tray icon to open the dashboard with stats, history, and actions. Right-click for the context menu.
With Razer Synapse (or similar software) you can map CTRL+ALT+D to a mouse button, e.g. Hypershift + scroll wheel click. Dictate without touching the keyboard.
Say the word, the tool inserts the character. The default mapping uses German words but can be customized in the SPOKEN_PUNCTUATION dictionary at the top of whisper-dictate.py.
| Spoken word | Result |
|---|---|
| Doppelpunkt | : |
| Semikolon | ; |
| Ausrufezeichen | ! |
| Fragezeichen | ? |
| Gedankenstrich | - |
| Schrägstrich / Slash | / |
| Anführungszeichen | " |
All settings are defined as variables at the top of whisper-dictate.py:
| Variable | Description | Default |
|---|---|---|
MODEL_SIZE |
Whisper model | large-v3 |
INITIAL_PROMPT |
Domain-specific terms for better recognition | Comma-separated list |
SPOKEN_PUNCTUATION |
Word-to-character mapping (regex) | See table above |
NO_SPEECH_THRESHOLD |
Silence detection threshold | None (disabled, VAD handles this) |
SHORT_TEXT_MAX_WORDS |
Remove trailing period for <= N words | 3 |
DEBUG_TRANSCRIPTION |
Write segment details to history log | True |
To switch the language, change the language="de" parameter in the model.transcribe() call to your language code (e.g. "en" for English).
Uses Whisper large-v3 with float16 precision and beam_size=5 for the best balance of quality and speed. Near-perfect accuracy for both English and German, including dialects and background music.
Benchmarks on RTX 4060:
| Scenario | Audio duration | Transcription time | Real-time factor |
|---|---|---|---|
| Short dictation (1-3 words) | 2-4s | ~0.5s | 4-6x |
| Medium dictation (1-2 sentences) | 4-10s | ~1s | 5-10x |
| Long dictation (6 sentences) | ~55s | ~5s | 11x |
| Very long dictation (20 segments) | 73s | 7.7s | 9.5x |
If you prefer faster transcriptions over maximum accuracy, switch to large-v3-turbo with beam_size=3 (~3-5x faster).
- Hotkey triggers audio recording via
sounddevice - Audio is captured as a NumPy array at 16kHz (no WAV file intermediary)
- Whisper transcribes with
faster-whisper(CTranslate2 backend) on your GPU - Post-processing applies spoken punctuation replacement and hallucination filtering
- Output is pasted into the active window via clipboard
The recording overlay uses pre-rendered animation frames (90 frames, 30fps) with 2D pixel displacement simulating SVG feDisplacementMap. A dual-ring system (inner plasma ring + outer orbit ring) with independent noise fields creates the electric border effect. All blur layers are pre-composited before the frame loop for minimal CPU usage during recording.
| Platform | Status | Reason |
|---|---|---|
| Windows 10/11 + NVIDIA GPU | Fully supported | Developed and tested |
| Linux + NVIDIA GPU | Not compatible | Uses Win32 APIs (kernel32, user32, winsound) |
| macOS (Intel/Apple Silicon) | Not compatible | No CUDA support, no Win32 APIs |
The Whisper engine itself (faster-whisper) runs cross-platform, but the integration layer (global hotkey, clipboard, overlay, system tray, audio feedback) is built on Windows APIs.
Contributions welcome! If you'd like to port Whisper Type to Linux or macOS, PRs are appreciated. The main components that need platform-specific replacements are:
- Global hotkey listener (
keyboardlibrary -> e.g.pynput) - Clipboard paste (
pyperclip+keyboard.send("ctrl+v")->xdotool/pbpaste) - System tray icon (
pystrayworks cross-platform, minor adjustments needed) - Recording overlay (tkinter with Win32 click-through -> platform-specific window flags)
- Audio feedback (
winsound.Beep-> e.g.simpleaudio) - GPU: Linux has CUDA support, macOS would need CoreML or CPU fallback
| Component | Minimum | Recommended |
|---|---|---|
| OS | Windows 10 | Windows 11 |
| GPU | NVIDIA with CUDA | RTX 3060+ |
| VRAM | 4 GB | 8 GB |
| Python | 3.12+ | 3.12+ |
| RAM | 8 GB | 16 GB |
Built by Daniel Gächter.
Check out my other projects:
- SEO Agent - SEO services and web development in Switzerland
- Lotus Academy - Tutoring school in German-speaking Switzerland
MIT