What are reliable approaches for TV audio classification, specifically for classifying whether incoming audio is a commercial?

Follow the full discussion on Reddit.
What I would like to do is sample television audio from a capture device, classify whether the audio is a TV commercial or not, and use that determination to respond in real time (at most a small number of seconds after a commercial break starts typically). Here is a project that does something similar. This project uses audio fingerprinting to detect whether a recorded audio fingerprint is found in a database of already identified commercials. The blog posts alludes to an alternate method of using machine learning to train a model to detect commercials that have never been observed before. That's what I'm reaching for.