Dictate

Speech-to-text dictation tool for Linux/Wayland, designed to run via keyboard shortcut. Press your hotkey to start recording, press again to stop. Your speech is transcribed locally using faster-whisper, optionally cleaned up by Claude, and copied to your clipboard.

Features

• Toggle-based recording: First invocation starts recording, second stops and processes
• Local transcription: Uses faster-whisper (Whisper implementation in CTranslate2) - no cloud API for transcription
• AI text cleanup: Optional grammar and punctuation cleanup via Claude CLI
• Direct typing: Optionally type text directly at cursor position (in addition to clipboard)
• Multi-language support: Transcribe in any language Whisper supports
• Desktop notifications: Progress updates via system notifications
• Configurable: Customize Whisper model, audio device, and prompt template

Requirements

System Dependencies

Install these via your package manager:

# Fedora
sudo dnf install portaudio-devel wl-clipboard libnotify

# For --type flag (optional)
sudo dnf install wtype

# Ubuntu/Debian
sudo apt install portaudio19-dev wl-clipboard libnotify-bin

# For --type flag (optional)
sudo apt install wtype

Claude CLI

The Claude CLI is required for text cleanup (unless using --raw mode):

# Install via npm
npm install -g @anthropic-ai/claude-code

# Authenticate
claude login

Python

Python 3.11+ is required. Using uv is recommended.

Installation

Clone the repository:

git clone https://github.com/yourusername/dictate.git
cd dictate

No additional installation steps needed - uv handles Python dependencies automatically.

Usage

This tool is designed to be triggered via a keyboard shortcut (see Setting Up a Keyboard Shortcut below). The same shortcut both starts and stops recording.

How the Toggle Works

1. First press: Starts recording (shows notification)
2. Second press: Stops recording → transcribes → processes → copies to clipboard

Testing from the Shell

When testing from a terminal, run in the background with & (otherwise Ctrl+C would kill the recording process):

# Start recording
uv run ./dictate.py &

# Stop and process (run again in any terminal)
uv run ./dictate.py

Command-Line Arguments

Argument	Short	Description
--language	-l	Language code for transcription (e.g., de, fr, es). Default: en
--raw	-r	Skip Claude processing, output raw transcription only
--type	-t	Type text at cursor position via wtype (in addition to clipboard)

Examples

# Transcribe in German
uv run ./dictate.py -l de

# Raw transcription without AI cleanup
uv run ./dictate.py --raw

# Type directly at cursor AND copy to clipboard
uv run ./dictate.py --type

# Combine flags
uv run ./dictate.py -l de -r -t

Setting Up a Keyboard Shortcut

Hyprland

Add to your hyprland.conf or keybinds config:

# Dictate - English (Super+Ctrl+Alt+E to toggle)
bind = $mainMod CTRL ALT, e, exec, uv run ~/code/dictate/dictate.py --type

# Dictate - German
bind = $mainMod CTRL ALT, d, exec, uv run ~/code/dictate/dictate.py --type -l de

# Raw mode (no Claude processing)
bind = $mainMod SHIFT CTRL ALT, e, exec, uv run ~/code/dictate/dictate.py --type --raw

Sway

Add to your ~/.config/sway/config:

# Dictate - English (Super+Ctrl+Alt+E to toggle)
bindsym $mod+Ctrl+Alt+e exec uv run ~/code/dictate/dictate.py --type

# Dictate - German
bindsym $mod+Ctrl+Alt+d exec uv run ~/code/dictate/dictate.py --type -l de

GNOME

1. Open Settings → Keyboard → Keyboard Shortcuts → View and Customize Shortcuts
2. Scroll to Custom Shortcuts and click +
3. Set:
   • Name: Dictate
   • Command: uv run /full/path/to/dictate.py --type
   • Shortcut: Your preferred key (e.g., Super+Ctrl+Alt+E)

Other Desktop Environments

Configure your DE's keybinding system to run the script. The same shortcut toggles recording on/off.

Configuration

Create or edit config.toml in the project directory:

[whisper]
model_size = "base"    # tiny, base, small, medium, large-v2, large-v3
device = "cpu"         # cpu or cuda
compute_type = "int8"  # int8, int16, float16, float32
language = "en"        # Default language

[audio]
# Uncomment to specify a specific audio input device index
# device = 0

[agent]
# Custom prompt template (uses {text} placeholder)
# prompt_template = """Your custom prompt here: {text}"""

Whisper Models

Model	Size	Speed	Accuracy
tiny	75 MB	Fastest	Lower
base	142 MB	Fast	Good
small	466 MB	Medium	Better
medium	1.5 GB	Slower	High
large-v3	3 GB	Slowest	Highest

For CUDA acceleration, set device = "cuda" and use compute_type = "float16".

How It Works

1. First run: Creates a PID file (/tmp/dictate.pid) and starts recording audio
2. Second run: Detects running instance via PID file, sends SIGUSR1 signal to stop
3. Processing pipeline:
   • Audio is resampled to 16kHz
   • faster-whisper transcribes the audio locally
   • Claude CLI cleans up grammar/punctuation (unless --raw)
   • Result is copied to clipboard (and typed if --type)
4. Notifications: Desktop notifications show progress at each stage

Troubleshooting

No audio input detected

List available audio devices and set the correct index in config.toml:

import pyaudio
p = pyaudio.PyAudio()
for i in range(p.get_device_count()):
    info = p.get_device_info_by_index(i)
    if info["maxInputChannels"] > 0:
        print(f"{i}: {info['name']}")

Claude CLI errors

Ensure you're authenticated: claude login

Use --raw mode to bypass Claude processing entirely.

CUDA not detected

Install CUDA-compatible versions:

uv pip install nvidia-cublas-cu12 nvidia-cudnn-cu12

License

MIT