Always-listening devices usually adopt a layered processing architecture: a low-power wake word detection module on the device, which activates the full audio capture and upload process only after detecting the trigger word.