I tend to think whatever good bits we see (Discovery being good) is sitting on the shoulders of that base work.
Yet, the product values of Spotify do not align with these people or musicians anymore.
Echo Nest is a set of music analysis routines that provide high level information about songs, like tempo, time signature, key, chordal modes, and other features best thought of as compositional/musical. Beneath it is a deep stack of low level signal processing parts, libaubio for example, for onset detection, spectral flux, timbre analysis, pitch estimation. These build into intermediates like chord analysis, rhythm classifiers and whatnot. Above that you can build "genre" and "mood" then cluster into a similarity space because there's few enough parameters.
It seems to me, more as a "mathematical intuition" at this stage, that advances in NN for generative work could be targeted at blind classification, bypassing the whole expensive parametric layers and do machine listening on the raw audio.
Are there any recommendation and music search systems that work this way. I expect the training prospect is terrifying, but whoever works it out first will have an alternative and possibly better route into music recommendation.
I'm not so sure they deserve all credit for Spotify's discovery features - those where at their best right after Sander Dieleman and Aäron van den Oord were interns there, and both those guys were snapped up by Deepmind and have a ridiculously good publication history by now.
But that Spotify is going to get a ton worse without MacDonald, is pretty certain.