oww-rs

A minimal, inference-only Rust port of openWakeWord. It reimplements the ONNX inference path — the shared melspectrogram → speech-embedding → wakeword-classifier pipeline — for low-latency wake-word detection that is easy to embed in a Rust application.

There is no training here: models are trained with the upstream Python project and consumed as .onnx files. The runtime is tract-onnx, so there is no native ONNX Runtime dependency to ship.

Features

Self-contained models — the melspectrogram + embedding front-end and the per-wakeword classifiers are embedded into the binary at compile time (rust-embed). Ships with alexa and hey_mycroft out of the box.
openWakeWord-compatible front-end — 16 kHz mono f32 audio in 1280-sample (80 ms) chunks, melspectrogram normalization and tensor shapes mirror the Python implementation so detection behaviour matches upstream.
Falling-edge detection smoothing — fires on the falling edge of the probability curve with a refractory window, which is more robust than a naive threshold crossing.
Microphone capture loop — a cpal-based capture/auto-reconnect loop (create_unlock_task_sync) that resamples any mic to 16 kHz and runs detection.
Built on audio_tools for mic config, resampling and chunking.

Quick start

Add the crate:

[dependencies]
oww-rs = "0.3"

Run inference over an audio stream, chunk by chunk:

use oww_rs::oww::{OwwModel, OWW_MODEL_CHUNK_SIZE};
use oww_rs::config::SpeechUnlockType::OpenWakeWordAlexa;

// threshold 0.1 — detection logic smooths over a ~1 s window
let mut model = OwwModel::new(OpenWakeWordAlexa, 0.1)?;

// Feed 16 kHz mono f32 chunks of OWW_MODEL_CHUNK_SIZE (1280) samples:
for chunk in audio_chunks {
    let detection = model.detection(chunk);
    if detection.detected {
        println!("wake word! probability = {}", detection.probability);
    }
}
# Ok::<(), Box<dyn std::error::Error>>(())

Live microphone demo

cargo run -p oww-rs --example cpal_test   # then say "Alexa"

The example (examples/cpal_test.rs) opens the default mic, resamples it to 16 kHz via audio_tools, and prints a line whenever the wake word fires.

Adding a new wake word

Train a classifier with the upstream openWakeWord project and export it to .onnx.
Drop the file into speech_models/.
Add a variant to SpeechUnlockType (config.rs) and wire it into OwwModel::new / new_model.

See the upstream openWakeWord project for pre-trained models and training instructions.

License