RU | EN

Wake word classifier training example

This directory contains a simplified, self-contained example of the training loop used for the wake word classifier bundled with the application:

16 kHz mono PCM WAV
-> openWakeWord-compatible ONNX feature extractor (from ../models)
-> small PyTorch DNN classifier
-> ONNX export in the format the application loads

It is an educational example, not the production training pipeline. The bundled models/privet_avrora*.onnx classifiers were trained on a much larger internal dataset that is not distributed with this repository.

Layout

• train_wakeword.py — the training script.
• requirements.txt — Python dependencies.
• demo_dataset/ — a tiny set of real recordings for demonstration (see demo_dataset/README.md); extend it with your own recordings to experiment.

Usage

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python3 train_wakeword.py

The script loads clips from demo_dataset/, extracts embeddings with the ONNX models from ../models, trains the classifier, prints validation metrics (TP/TN/FP/FN, recall), and writes the result to outputs/:

• wakeword_classifier.onnx — classifier in the same format as ../models/privet_avrora.onnx (input features, output score).
• wakeword_classifier_metadata.json — shapes and validation metrics.

Useful options:

python3 train_wakeword.py --epochs 50 --batch-size 8 --layer-size 64 \
    --dataset /path/to/my_dataset --output-dir /path/to/outputs

Notes

• CPU is enough; the dataset and the model are tiny.
• The feature extraction pipeline in train_wakeword.py mirrors the C++ code in ../src/wakeworddetector.cpp step by step, so a classifier exported here can be dropped into the application next to (or instead of) the bundled one.