oww-rs — minimalistic OpenWakeWord inference in Rust

This is extracted source code from openWakeWord, providing simplified code for inference of ONNX models only. The Python inference path is converted to Rust for better wake-word detection performance. Detection runs on the tract-onnx runtime, supporting all major platforms. The alexa.onnx model is bundled as a sample — simply saying "Alexa" into the microphone triggers a detection, as implemented in crates/oww/examples/cpal_test.rs.

Microphone access uses the pure-Rust cpal crate, which works on all major platforms.

See the original openWakeWord project for pre-trained models and for training new custom wake words. Training is not included here — use openWakeWord's steps to generate a model for a custom wake word.

Workspace layout

This is a Cargo workspace with two publishable crates:

Crate	Published as	What it is
crates/audio_tools	audio_tools	Reusable, model-agnostic mic capture, resampling, channel handling, conversion, RMS and WAV saving for a 16 kHz mono pipeline.
crates/oww	oww-rs	The wake-word ONNX inference (tract) + mic-capture detection loop. Depends on audio_tools.

Running

cargo run -p oww-rs --example cpal_test            # live mic demo — say "Alexa"
cargo run -p audio_tools --example mic_to_chunks   # mic → 16 kHz chunk demo
cargo test                                         # run the whole workspace

Contribution

Open an issue / PR.

License

MIT license, see LICENSE