Single pixel imaging is a well known method that depends on injecting known information (light patterns) and capturing the same scene multiple times in order to reconstruct.

Audio is by definition a time varying signal. If you cannot sample it fast enough, the information is gone. The fundamental Nyquist-Shannon sampling theorem has proven the concept of the critical (Nyquist) frequency. There are ways around this in special cases: for band limited signals undersampling is a method to reconstruct them using sampling lower than the critical frequency, and for sparse signals compressed sensing can be used. Real life audio is neither band limited (as sampled by any device on the phone), nor sparse. I think it is a physical impossibility with the current sensors.

https://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampli...

To reiterate, I'm not suggesting it should be able to listen in on conversations, but rather that it may plausibly be able to detect recurring environmental vibrations. For example, let's say your phone is left on a table unnattended, then if your house is near a train then it may be able to sense the regularity of the train's passing (whose schedule itself varies throughout the day, providing a unique fingerprint for both determining which train it might be and your position along that train's route).