> Or was this going to upload the audio to Google's servers to do the work? If it's the latter, that would necessarily[1] require uploading audio without a wake

> Or was this going to upload the audio to Google's servers to do the work? If it's the latter, that would necessarily[1] require uploading audio without a wake-word trigger. Couldn’t it run a local model to detect possible incidents, and when a local confidence threshold was exceeded, upload to Google to run a more intense model? I’m pretty sure this is how things like “Hey Siri” and “okay Google” are implemented.