Indeed. I see two options here.

First is to analyze the other signal features of the commercials (eg. increased volume), although it may be tricky.

The other option is a crowd-sourced solution - pretty much as for the browser adblock - where users can mark samples recognized as ads. Since the publishers often buy campaigns for many stations in the same country or state, it may be a shared database.

On the other hand, the described project only scratches my own itch. I wouldn't try to productise an app that takes away the main source of income for the radio stations.

At least in Germany, they all have. Also you could almost time it, since commercials start ~5mins before the full hour (where the news comes on).

I've often toyed with the same idea of the parent poster

I wonder, if you accepted a little delay, say 20 seconds, could you use Shazam to identify what song is currently playing? I have been amazed at the range of music Shazam can identify, if you could do regularly queries and separately listen out for gaps in music, it seems plausible to be able to identify when a non-song (from Shazam's perspective) is playing.