Source: Apple Machine Learning Journal
Possibly the most interesting thing about these articles is their source. Apple really isn’t know for candour regarding the implementation of it’s products. That’s still somewhat in evidence here, as this doesn’t have the kind of rigorous description which would allow you to duplicate their results. That said, there are some interesting titbits here, particularly the way the sound input is broken down into something which can be fed into a neural net, and the multi-stage approach used to reduce the number of false alarms.