Whisper Dictation

Local GPU speech-to-text dictation tool. Hold a hotkey to record, release to transcribe and type the result into the active window. Runs fully offline — no cloud, no API key.

Features

• System tray icon with settings GUI (tkinter)
• Configurable hotkey, model, language, audio device
• Cross-platform: Windows and Linux builds from a single codebase
• Shared config via git (config.json, vocabulary.json)
• Machine-specific settings stored locally (audio device, GPU settings, model)
• Configurable shared paths for vocabulary and model cache (useful for dual-boot setups)

Requirements

Windows

• Python 3.13
• NVIDIA GPU with CUDA 12 drivers
• PortAudio (bundled with most Python sounddevice wheels)
• pywin32 (for system tray and keyboard injection)
• pyinstaller (for building a standalone executable)

Linux

System packages (install via package manager):

Arch/CachyOS:

sudo pacman -S tk libayatana-appindicator wl-clipboard xdotool

Debian/Ubuntu:

sudo apt install python3-tk libayatana-appindicator3-1 wl-clipboard xdotool

Package	Purpose
tk	tkinter GUI (settings, log, vocabulary windows)
libayatana-appindicator	System tray icon (required for KDE/GNOME on Wayland)
wl-clipboard	Text injection on Wayland (wl-copy)
xdotool	Simulates Ctrl+V paste on Wayland, text typing on X11

Optional (for GPU acceleration):

Arch/CachyOS:

sudo pacman -S nvidia cuda

Without CUDA, the app runs on CPU. Use int8 compute type and a smaller model (small or base) for acceptable speed on CPU.

Python:

• Python 3.10+
• PortAudio (bundled with sounddevice wheels)

Installation

Windows

install.bat

This creates a .venv-windows virtual environment, installs all dependencies and the CUDA 12 DLLs required by faster-whisper.

Linux

chmod +x install.sh start.sh build-linux.sh
./install.sh

Creates a .venv-linux virtual environment with all dependencies and PyInstaller.

Usage

Windows

start.bat

Linux

./start.sh

The app starts in the system tray. Hold the hotkey (default: Ctrl+Shift+Space) to record, release to transcribe and type into the active window.

Build

Builds are platform-specific and output to separate directories:

• Windows: dist/whisper-dictation-windows/
• Linux: dist/whisper-dictation-linux/

Windows

.venv-windows\Scripts\python.exe build.py

Linux

./build-linux.sh

Both use PyInstaller to bundle the app into a standalone folder. The resulting executable can be run without a Python installation.

Configuration

Shared config (config.json, in app directory)

Key	Default	Description
hotkey	ctrl+shift+space	Recording trigger
language	de	Transcription language (de, en, fr, es, it, null = auto)
sample_rate	16000	Audio sample rate in Hz
vocab_path	""	Path to vocabulary file (empty = local vocabulary.json)
model_dir	""	Path to shared model cache directory (empty = default HuggingFace cache)

Local config (config_local.json, per machine)

Stored outside the app directory to keep machine-specific settings separate:

• Windows: %LOCALAPPDATA%\WhisperDictation\config_local.json
• Linux: ~/.local/share/WhisperDictation/config_local.json

Key	Default	Description
model	medium	Whisper model size (tiny, base, small, medium, large-v2, large-v3)
device	cuda	Inference device (cuda or cpu)
compute_type	float16	Precision (float16 for GPU, int8 for CPU, float32)
audio_device	null	Microphone (null = system default)

Sharing data between Windows and Linux

On a shared drive (e.g. Ventoy USB), both builds can use the same vocabulary and model files. Set vocab_path and model_dir in the Settings UI to point to a common directory:

shared_data/
  vocabulary.json    <- shared vocabulary
  models/            <- shared Whisper model cache

Audio settings, model selection, and compute type remain per-platform in config_local.json.

Vocabulary

Custom vocabulary/replacements can be edited via the Settings UI or directly in vocabulary.json. Words are passed as initial prompts to improve recognition of domain-specific terms. Replacements are applied as find/replace after transcription.

Model Download

On first start the selected Whisper model is downloaded automatically from HuggingFace (~500 MB for medium). Subsequent starts use the cached model. Set model_dir to share the cache between builds.