If you've ever been to a sports bar during a big game, you'll notice often they play music during commercial breaks so people don't have to listen to commercials. Almost always it's a human being manually switching the audio feed. Sometimes it's one of the bartenders, other times there is a hired DJ for this sole purpose.
With the help of a neural network engineer I built a prototype on a Raspberry Pi 3 that is quite effective at automating this task. The program is trained with 80 gb of labelled audio data from TV commercials and sporting events (this data is available as well). Using audio only, the program analyzes incoming data from a live feed, making a decision on when to switch audio output based on a minimum confidence threshold that a commercial break has started or ended.