ghosttype

Hold-to-talk voice typing for Windows. Press a hotkey, speak, release — your words appear in whatever app has focus. Fully offline. Local Whisper.

This is Greg DeYoung's fork of CatBoneheaD/Whisper-Writer, maintained for personal use across the Greg DeYoung fleet.

What it does

Hold a hotkey, speak, release. Your voice is transcribed locally by faster-whisper and typed into the focused window — same workflow as BoxType/Voxtype on Omarchy Linux. Recording modes include hold_to_record (BoxType-style push-to-talk), press_to_toggle, continuous, and voice_activity_detection.

Features

• Modern dashboard (Voice Ink-style) with dark & light themes — sidebar with History / Settings, live status indicator.
• Transcription history stored locally in SQLite — search, one-click copy, re-insert into the active window, delete.
• In-app model switcher — full multilingual set (tiny…large-v3); not-installed models download on selection.
• Language picker & translation — choose spoken language or translate to English.
• Text post-processing — spoken-command replacements, auto-capitalization, trailing space.
• Settings UI — hotkey with press-to-capture, recording mode, language, device, theme, autostart — no YAML required.
• Start with Windows (optional) and silent launch via launch.vbs — lives in the system tray.
• Single-instance guard — no more doubled text from a second launch.
• Hold-to-record — BoxType-equivalent push-to-talk mode.

Transcription runs locally by default (faster-whisper), or through the OpenAI API if you enable it in Settings.

Pedigree

ghosttype is built on the work of three upstream projects. All credit goes to the original authors.

ghosttype (this fork)
    │
    └── CatBoneheaD/Whisper-Writer  (active 2026, Windows-first fork)
            https://github.com/CatBoneheaD/Whisper-Writer
            Added: modern dashboard, hold_to_record mode, in-app model
            switcher, transcription history, system tray, single-instance
            guard, Windows install fixes (av==11.0.0 workarounds),
            press-to-capture hotkey, light theme, post-processing.

            └── savbell/whisper-writer  (original project, 1.1k★, MIT)
                    https://github.com/savbell/whisper-writer
                    The original WhisperWriter — PyQt dashboard,
                    faster-whisper integration, push-to-talk, evdev
                    input layer, recording mode framework.

                    └── openai/whisper  (the model, not a code fork)
                            https://github.com/openai/whisper
                            The underlying speech recognition model.
                            ghosttype uses the faster-whisper inference
                            runtime against OpenAI's Whisper weights.
                            No code is derived from OpenAI's repository;
                            we depend on the trained model artifacts only.

The license chain breaks at CatBoneheaD's relicense in their 2026 rewrite (MIT → GPLv3), because the Qt dashboard work they added is GPL-compatible.

Getting started

Prerequisites

• Git
• Python 3.11
• (Optional, for GPU) an NVIDIA GPU. The required nvidia-cublas-cu12 and nvidia-cudnn-cu12 libraries install automatically via requirements.txt, and the app adds them to PATH at startup — no manual CUDA setup needed.

Installation

# 1. Clone
git clone https://github.com/gdeyoung/ghosttype.git
cd ghosttype

# 2. Create and activate a virtual environment
python -m venv venv
venv\Scripts\activate        # Windows

# 3. Install dependencies
pip install -r requirements.txt

# 4. (Optional) Pre-download a local model
python download_model.py

# 5. Run
python run.py

Launch without a console (recommended)

Double-click launch.vbs — it starts the app silently with pythonw.exe (no console window). Create a desktop shortcut to launch.vbs and set its icon to assets/ww-logo.ico.

Default hotkey is Ctrl+Shift+Space. Change it in Settings — the press-to-capture button makes it easy to set F9 to match your Omarchy BoxType workflow.

Recording modes

• continuous (default): stops after a pause, transcribes, then keeps listening. Press the hotkey again to stop.
• voice_activity_detection: stops after a pause; won't restart until you press the hotkey.
• press_to_toggle: starts on hotkey, stops on the next hotkey press.
• hold_to_record: records while the hotkey is held down. Set this to match BoxType.

Configuration

Everything can be set in the in-app Settings page. Settings save to config.yaml (git-ignored, your personal config stays local). Missing values fall back to defaults in src/config_schema.yaml.

License

GPLv3 — see LICENSE. Inherited from CatBoneheaD's fork. Any derivative work must also be GPLv3.