Minimalistic version of the OpenWakeWord inference in Rust and ONNX runtime

This is extracted source code from the OpenWakeWord providing simplified code for inference of onnx models only. The Python code for inference is converted to Rust for better wake word detection performance. The inference uses ORT supporting all major platforms. For detection the hey_jarvis_v0.1.onnx model is used as a sample from the OpenWakeWord project. There is also hey_jarvis.wav file used in test to trigger wake word detection with more than 99% accuracy.

The implementation uses pv_recorded opensourced audio recorder from Picovoice. The recorded uses underlying miniaudio lib for accessing microphones on all major platforms.

See the original OpenWakeWord lib for more details for pre-trained models or training new custom model. The training stage is not included in this repo and the OpenWakeWord's original steps should be used to generate new model triggering on a custom wake word.

If you are looking for simplified Python version, see git log with tag python_version.

Running like this

cargo run --example from_mic

or tokio based version

cargo run --example from_mic_tokio

Running tests:

cargo test

Contribution

Make an issue / PR etc.

Licence

MIT license, see LICENSE