pantorn/wuwei-hey-wu

Wuwei — Hey Wu Wake Word Model

Trained openWakeWord model for the wake phrase "Hey Wu".

Model

• Format: ONNX (13KB)
• Architecture: 2-layer DNN (128 units) on 96-dim audio features
• Training: 10000 steps + model selection via openWakeWord pipeline
• Dataset: 120 positive clips (espeak-ng TTS) + 16GB ACAV100M negative features
• Framework: openWakeWord + PyTorch + Docker (Linux)

Usage

import openwakeword
model = openwakeword.Model(
    custom_model_paths={"hey_wu": "hey_wu.onnx"}
)
# Feed 1280-sample int16 chunks
result = model.predict(chunk)
score = result.get("hey_wu", 0.0)
# If score > 0.7, wake word detected

Training

See train_hey_wu_colab.py for the full training pipeline.

Quick retrain (Docker)

docker build -f Dockerfile.train -t wuwei-train-img .
docker run --rm -v /path/to/cache:/root/wuwei-training-data wuwei-train-img

Quick retrain (Google Colab CLI)

colab run --gpu T4 train_hey_wu_colab.py

License

Apache 2.0 (openWakeWord). Model trained on ACAV100M features (CC BY 4.0) + synthetic TTS clips.