I read up about it a while back, so anyone pleas correct me. It works like this I think: An audio sample is recorded, chunked into overlapping windows, which are converted into a spectagram (the black, blue, yellow, red image of frequency vs time). Idk at which point what is send to the server but to me it would make sense to keep the data amount low by sending after the next step: From the spectagram, the algorithm searches for high spots or notable “pixels” in the spectagram. It then notes the relative position between many spots, not only the next but lets say 10 per spot (propably more, as just a pair of 2 numbers are super cheap and small). We then hash these relative vectors. So we now have a list of frequency vs time vectors hashed in unique strings of our recorded audio windows. In the database we have all songs send through the same algorithm and stored. We then just compare our hashes from the recording to the database and find the song with the most matches.

Its pretty smart and I suppose there are additional filters to speed up the process and make it more reliable. Its just so simple, I love it