I have a feeling they are using federated machine learning for this to have a lot of the processing done locally and not need to activate the radios for as much of the processing as possible. They have been making big strides in that area lately and this might be the start of some of the major applications of it (I think they are using it in their keyboard prediction as well from a bit ago)
anyone know how many gb or mb this occupies in my phone... can v jus clear this data...
More likely (this is a guess, nobody outside google really knows at this point), they will use federated machine learning to figure out that something is "a song", then perhaps clean up and isolate the actual "song" part of it and send that over to a google server for processing.
But again it was just announced, so nobody really knows how this works, where the data is or goes, and what tradeoffs were made.
[edit] typos
I'm probably underestimating the parameters per node, and overestimating the size of the layers closer to the input. Further, it's more likely structured as an LSTM than a convolutional network, since sound is a streaming source.