Wake word
Threshold
A higher threshold means fewer false activations and fewer detections. On a calibrated model, 0.5 targets about one false activation per hour on the calibration audio.
Audio
Score
Loading the model…
About
These are the ready wake-word models of ovos-ww-plugin-wakeforge, the OpenVoiceOS wake-word plugin, published in OpenVoiceOS/wakehubert-wakewords. Each model is a small classifier on features from WakeHuBERT tiny, a 0.64M-parameter speech feature extractor distilled from HuBERT-base, and the models were trained with wakeforge.
The page scores audio the way the plugin does: every 80 ms it featurizes the last 1.5 s of audio on its own, runs the wake-word model on the features and takes the sigmoid of its output as the score. A score at or above the threshold is a detection, after which the model is reset and detections are ignored for 2 s. The models run in onnxruntime-web.
Recall and false activations per hour are the figures the model repository publishes for each model at its default threshold; the page reads the model list, the models and these figures from that repository on every load. The example clips are synthetic speech from the plugin's test clips; any clip can be scored by any model, which shows how similar-sounding words behave.