Only 53 of 1,045 Hours of Sleep Audio Were Usable for Training a Sleep-Phase Model

A developer team recorded about 1,045 hours of overnight bedroom audio to train a sleep-phase model, but only roughly 53 hours met the criteria for training. The main obstacle was not capturing audio, which they described as a solved problem, but obtaining reliable labels that mark sleep stages. Their labels came from wearable devices, which only worked on nights when the phone and wearable recordings overlapped cleanly, and small clock drift or early device failures removed usable hours. They concluded that defining what counts as a valid night, and discarding incomplete recordings, was more important than collecting more data. Their practices included logging each recording's completeness, aligning audio to labels before labelling, and tracking where each label came from.