Over the last year the tech product that has brought the most joy into my life has undoubtedly been Discover Weekly, Spotify’s playlist of tracks personalized for my taste. Today the company is rolling out another feature along the same lines: Release Radar gives you a weekly playlist of songs culled from new albums. Discover Weekly will kick things off each Monday, and Radar will take you into the weekend with a Friday release. It’s actually a much more challenging task to get right, because unlike the tracks from Discover Weekly, there isn’t a lot of good data available on brand new music.
“When a new album drops, we don’t really have much information about it yet, so we don’t have any streaming data or playlisting data, and those are pretty much the two major components that make Discover Weekly work so well,” says Edward Newett, the engineering manager at Spotify in charge of Release Radar. “So some of the innovation happening now for the product is around audio research. We have an audio research team in New York that’s been experimenting with a lot of the newer deep learning techniques where we’re not looking at playlisting and collaborative filtering of users, but instead we’re looking at the actual audio itself.”
Release Radar takes stock of your entire listening history
Discover Weekly focuses on a window of the last six months or so to decipher your taste and make suggestions. Release Radar can’t replicate that approach, because your favorite band may not release an album more than once every two years. Instead it takes stock of your entire listening history, then narrows the range of possible suggestions down to tracks that have been released in the last two to three weeks.
Release Radar is trying to do more than just highlight new music from acts it knows you like. But suggesting new bands can be hard, because brand new music often hasn’t been categorized yet. “How are we sure in fact that this is Covenant the Swedish pop band and not Covenant the Norwegian black metal band?” asks Newett. In the past, if two artists had the same name and were injected into the same catalog, Spotify needed users to go through and annotate it manually. Luckily, deep-learning algorithms are now capable of doing this kind of assessment quickly at at massive scale.
In a way, this technique brings curation back full-circle to the approach Pandora began employing over a decade ago, when it hired musicologists to listen to tracks and assign them certain characteristics. “I think it’s a little bit like what Pandora was doing,” says Newett. “I think we’re more on the cutting edge of what’s this new deep learning is trying to tackle. It would basically be automated ways of figuring out what music is similar to other music, looking at all the different audio signatures of the song.”