Why would random snapshots of audio from a watch be useful AI training data?