A privacy-focused voice command system for developers that works entirely offline. Uses OpenAI's Whisper AI locally on your machine for speech-to-text and WebRTC VAD for voice activity detection.
Copyright 2025 Adrian Scott
- 100% Offline - No audio data leaves your machine
- Real-time Monitoring - Continuous voice input detection
- IDE Integration - Direct text insertion into code editors
- Voice Commands - Custom commands for common actions
- Command Word Support - "Flow" prefix for commands (configurable)
- Pause/Unpause Transcription - Say "Flow pause" or "Flow unpause"
- GPU Acceleration - Optional CUDA support for faster processing
- Ubuntu 20.04+ (other Linux distros may work)
- Python 3.8+
- Working microphone or audio input device
- xdotool (
sudo apt install xdotoolon Ubuntu/Debian) - PortAudio libraries (
sudo apt install portaudio19-dev)
- Clone the repository
- Install dependencies:
sudo apt install portaudio19-dev python3-dev xdotool ffmpeg pip install -r requirements.txt
Froshine can be configured using either:
- Command-line arguments
- Environment variables in a
.envfile - System environment variables
Command-line arguments take precedence over environment variables.
Copy the example configuration file to create your own:
cp .env.example .envThen edit .env to customize your settings. See .env.example for available options.
By default, Froshine uses your system's default audio input device. You can configure the audio input using environment variables:
FROSHINE_AUDIO_DEVICE: Specify a preferred audio device by name or indexFROSHINE_LIST_DEVICES: Set to "1" to list all available audio devices
Examples:
# List all available audio devices
FROSHINE_LIST_DEVICES=1 python voice_monitor_command_word.py
# Use a specific device by name (partial match)
FROSHINE_AUDIO_DEVICE="USB" python voice_monitor_command_word.py
# Use a specific device by index
FROSHINE_AUDIO_DEVICE="2" python voice_monitor_command_word.pyFroshine supports different Whisper models for speech recognition. You can choose the model using the --model or -m flag:
# Use the tiny English model (fastest, less accurate)
python voice_monitor_command_word.py --model tiny.en
# Use the large v3 model (slower, most accurate)
python voice_monitor_command_word.py --model large-v3Available models:
tiny.en: Tiny model (English only) - Fastest, lowest accuracybase.en: Base model (English only) - Fast, basic accuracysmall.en: Small model (English only) - Default, good balancemedium.en: Medium model (English only) - Better accuracy, slowerlarge-v3: Large v3 model (All languages) - Best accuracy, slowest
The default model is small.en, which provides a good balance between speed and accuracy.
This script continuously listens for voice input, transcribes it locally with Whisper, and types the transcribed text directly into your active window.
Start the script:
python3 voice_monitor_command_word.pyBegin speaking: The system will detect speech and automatically type the transcribed text into your currently focused application.
Issue commands:
- Say "Flow enter" to press Enter.
- Say "Flow save file" to simulate Ctrl+S.
- Say "Flow pause" to stop typing text (commands still work).
- Say "Flow unpause" to resume typing text.
- Stop the script: Say "Flow quit", or press Ctrl+C in the terminal to exit.
Common Issues:
- ALSA/JACK warnings: Normal and safe to ignore
- No audio input:
# Check recording devices arecord -l - Permission issues:
sudo usermod -a -G audio $USER # Reboot after running
- All audio processing happens locally
- No internet connection required
- No tracking or data collection
Copyright 2025 Adrian Scott
Acknowledgements:
- OpenAI Whisper AI model
- WebRTC VAD for voice detection
- PyAudio for audio capture
early work in progress
```bash
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
sudo apt install xdotool
WINDOW_ID=$(xdotool search --name "\(Workspace\) \- Windsurf")
echo $WINDOW_ID
xdotool windowactivate --sync $WINDOW_ID; xdotool type --window $WINDOW_ID --delay 0 "windsurf test froshine"
Current mechanism is to start voice recorder, voice_to_ide.sh, then click in the field of Windsurf I want it to go into.
Next step: voice detection to automatically fire up the recorder.
After that: voice commands to choose window, and especially use Freepoprompt and o1-xml-parser to update files.