⚡ I started this project because I couldn't find a good open-source tool for wake word detection — most commercial solutions are expensive, and for my future voice assistant app, it makes more sense to build the functionality in-house.
This repository is designed to help train a neural network for Wake Word Detection (WWD) and integrate it into your own application.
It includes:
- Tools for recording and preprocessing audio data
- Scripts for training, testing, and buffered inference
- Future support for data augmentation (noise overlay, time stretching, compression, etc.)
We use a lightweight LSTM-based neural network, ideal for real-time, low-resource applications.
Training is performed inside a Jupyter Notebook, with step-by-step explanations.
- Data augmentation pipeline (noise overlay, tempo shift, pitch shift, etc.)
- Dataset balancing and visualization tools
To train the model, we need labeled wake word and background samples.
A simple recording script is included:
scripts/prepare_data.py— collect wake words and background noise- Files are saved as
.wavin structured folders
The file scripts/run_detector.py shows how to use the trained model in a live setting with a sliding audio buffer.
This module:
- Continuously records audio in short fragments (e.g., 0.25s)
- Maintains a buffer of recent fragments
- Triggers detection when the buffer is full
Useful for integrating into a voice assistant frontend.
src/
├── buffer/ # Sliding audio buffer implementation
├── model/ # Wake word detection model wrapper
├── recording/ # Audio recording tools
scripts/ # CLI scripts for training / running
tests/ # Unit tests
saved_model/ # Trained models (excluded from repo)
uv pip install -r pyproject.tomlUse the test scripts and pytest to verify logic inside the buffer and prediction flow.
You can also feed in your own audio segments to test individual cases.
MIT — Free for personal and commercial use.
Built with ❤️ by an engineer who wants WWD to be simple, cheap, and effective.