Whisper Dictation

Privacy-first local speech-to-text for NixOS -- whisper.cpp powered, push-to-talk, paste anywhere.

Features

• 100% local and private -- no cloud, no telemetry, works fully offline
• Push-to-talk -- hold Super+Period, speak, release to paste text
• Real-time feedback -- floating GTK4 window shows transcription status
• Multilingual -- supports 99 languages with auto-detection
• Wayland native -- built for GNOME on Wayland, works in any application
• Optimized for technical speech -- tuned for developer and AI workflows

Requirements

• NixOS or any Linux distribution with Nix
• Wayland compositor (GNOME recommended)
• PulseAudio or PipeWire
• User must be in the input group for keyboard monitoring

Installation

NixOS (recommended)

Add to your flake.nix:

{
  inputs.whisper-dictation.url = "github:jacopone/whisper-dictation";

  # In your configuration
  environment.systemPackages = [
    inputs.whisper-dictation.packages.${system}.default
  ];

  # Enable auto-start
  systemd.user.services.whisper-dictation = {
    enable = true;
    wantedBy = [ "graphical-session.target" ];
  };
}

Manual

git clone https://github.com/jacopone/whisper-dictation.git
cd whisper-dictation
nix develop
python -m whisper_dictation.daemon

First-time setup: ensure your user is in the input group (sudo usermod -aG input $USER, then log out and back in), download a Whisper model to ~/.local/share/whisper-models/, and start the ydotoold daemon. See the first-time setup section in DEVELOPMENT.md for details.

Usage

Start the daemon and dictate:

run-daemon          # use config file settings
run-daemon-en       # English only (fastest)
run-daemon-it       # Italian only
run-daemon-auto     # auto-detect language (adds ~1-2s)

Then in any application:

1. Click in a text field
2. Hold Super+Period
3. Speak naturally
4. Release the key -- text is pasted instantly

Override settings per-session with command-line flags:

python -m whisper_dictation.daemon --verbose --language auto --model base

Configuration

Edit ~/.config/whisper-dictation/config.yaml. Key settings:

• whisper.model -- model size: tiny, base (recommended), small, medium, large
• whisper.language -- language code (en, it, auto, etc.)
• whisper.use_gpu -- use the GPU when whisper.cpp is built with GPU support (see GPU acceleration)
• hotkey.key / hotkey.modifiers -- push-to-talk keybinding

See config.yaml in the repository for all available options with documentation.

Model selection guide

Model	Size	Speed	Accuracy	Use Case
tiny	39 MB	~1-2s	60%	Quick notes, testing
base	142 MB	~4-6s	70%	Recommended for speed
small	466 MB	~10-15s	80%	Balanced performance
medium	1.5 GB	~20-30s	85%	High accuracy
large	2.9 GB	~40-60s	90%	Maximum accuracy

Times measured on CPU (4 threads). GPU acceleration can reduce times by 5-10x.

GPU Acceleration

Transcription runs through the whisper-cli binary from whisper.cpp, so GPU acceleration is determined entirely by how that binary is built. The default package is CPU-only.

Vulkan (recommended -- works on AMD, Intel, and NVIDIA GPUs):

environment.systemPackages = [
  inputs.whisper-dictation.packages.${system}.whisper-dictation-vulkan
];

Or try it directly:

nix run github:jacopone/whisper-dictation#whisper-dictation-vulkan

CUDA or ROCm: build against your own whisper.cpp using the exposed builder (CUDA requires nixpkgs.config.allowUnfree = true):

environment.systemPackages = [
  (inputs.whisper-dictation.lib.${system}.mkWhisperDictation
    (pkgs.whisper-cpp.override { cudaSupport = true; }))   # or rocmSupport
];

When whisper-cli is built with GPU support it uses the GPU automatically -- no configuration needed. To force CPU on a GPU build, set whisper.use_gpu: false in config.yaml.

How It Works

1. Keyboard monitoring -- evdev captures low-level key events
2. Audio recording -- ffmpeg records microphone input while the key is held
3. Transcription -- whisper.cpp processes audio locally on your machine
4. Text insertion -- ydotool pastes transcribed text into the active window
5. UI feedback -- GTK4 floating window shows real-time status

Comparison

Feature	Whisper Dictation	Aqua Voice	Talon Voice
Privacy	Local	Cloud	Local
Cost	Free	$8/mo	$15/mo
NixOS support	Native	No	Manual
Technical terms	65-85%	97%	95%
Wayland	Yes	Limited	X11 only
Real-time	Yes	Yes	Yes

Development

See DEVELOPMENT.md for the full development guide.

Troubleshooting

See TROUBLESHOOTING.md for solutions to common issues (audio, keyboard detection, ydotool, hotkeys, performance).

Contributing

Contributions welcome. See CONTRIBUTING.md for guidelines.

License

MIT License -- see LICENSE.