It is pretty neat and several resources online explaining it but it chunks the recoding and does a hash of the fft of each frame i think which reduces the data a lot. The server has a pre hashed database https://hackernoon.com/how-shazam-works-in-a-nutshell