100DB was a music discovery project I started. This is part of a series of posts I wrote to explore the thinking behind it.
There are a lot of music platforms out there. How good are they at ‘music discovery’? What does that even mean?
Spotify’s “Discover Weekly” playlist seems to be heavily influenced by whatever I listened to last. Caveat: I don’t follow friends or have any social media accounts connected, so I’m mostly looking at the output of an algorithm that churns up ‘similar artists’. Apparently the last time I was on Spotify I was blasting non-stop trap music (the silliest genre of electronic music) because my algorithmic playlist is chock-a-block full of nothing but hilarious dirty drops right now.
In terms of album recommendations: lots of artists I’ve never heard of (great!) that all turned out to be more trap music (not necessarily great! Challenge me, Spotify!). There was also a ‘because you listened to’ section, which turned up some more interesting suggestions. My uninformed presumption is that some kind of node/edge math is happening behind the scenes to find things that fit a certain kind of musical profile. There was a lot more variation in sound here, and some albums I genuinely want to explore more deeply.
Exposing network-effects connections definitely yields more interesting picks than just ‘the same genre’, which is not surprising. Figuring out how to dial in the ‘edge length’ for uniqueness vs. likeability is probably someone’s dissertation.
While I was researching, I found this post, a helpful and deeply technical dive into how Spotify actually works their magic which I’ll no doubt revisit in the future.
I’ve bought the most music through Bandcamp, so they should have a fairly solid profile on me with which to make recommendations. They have a personalized feed I can explore, which keeps me up to date on new releases from artists I’ve supported previously (sweet!) but doesn’t do a heck of a lot to turn me onto other artists that fall into the locus of my listening interests.
The interesting thing it does do is suggest ‘fans’ to follow, who presumably(?) have similar tastes to me. This is an extremely Twitter-esque approach, which I like – I think there’s a really interesting pivot point between intentionality and algorithm that 100db could occupy. Jumping into a fan’s profile gives me a lot of datapoints but not a lot of direction or priority. I can’t, at first blush, understand why this fan might have been suggested to me as interesting, other than broad strokes of similar taste. I should note that albums in a fan’s ‘collection’ are also tagged with how often they appear in other collections globally, which is an interesting proxy for popularity but doesn’t do much to suggest how much I personally will enjoy it.
There is no other mechanism to suggest music to me, even though would be in Bandcamp’s interest to do so. If you read that Medium link above, you know that Spotify’s poured a truly bonkers amount of human and computer time into recommendations, which just may not be possible for Bandcamp. Doing a quick’n’dirty headcount on their ‘About’ page suggests about 50 employees – Spotify, in comparison has more than 4,000. Also extremely notable: Bandcamp is a privately held company, whereas Spotify has never been profitable and has raised nearly 3 billion dollars over its lifetime. The fact that I’m talking about Bandcamp and Spotify in the same breath is actually buckwild, on that basis.
Soundcloud doesn’t appear to have any recommendation features, but does have charts and ‘moods’, both of which feel like table stakes. I tend to only follow DJs on Soundcloud, which definitely gives me a very myopic view of what Soundcloud is ‘for’.
Soundcloud has a ‘repost’ feature which is like Twitter’s retweet, which gives it a credible method to build ‘feeds’ that I might want to follow. It also has pretty robust tagging and deep search, which is how I’ve discovered the music I’ve enjoyed the most on their platform. Because I’m typically coming to Soundcloud for DJ mixes, being able to search by the ‘physical’ qualities of music – length being the important quality in this instance – is a key feature.
My hot take is that Soundcloud, to make a terrible comparison, is the Behance of music: it’s as much about posting music as it is listening to it – a place to build a profile, to be seen. There’s something really interesting about that, although I’m not sure (yet!) what that means in terms of music discovery.
Youtube’s a weird one. I actually probably listen to the MOST music through Youtube – I can pull up virtually any song (or alternate version, or remix, etc.) and my expectation is that they’ll have everything. It’s also a very performative platform in a way that the rest aren’t – if I want to share nearly any piece of media (whether by sending a link or in real life), Youtube is going to be my go-to. It has ubiquity of presence and availability in a way that literally no other platform has.
Youtube also has ‘Youtube Music’, which feels neutered and strangely empty compared to the overturned-beehive-overload of Youtube proper. I can’t tell if it’s recommending anything for me or not, but given that the second ‘shelf’ of suggestions on the site is “Music For Hockey Night”, I have to assume that it knows literally nothing about me.
I will give Youtube Music props for the deep links it’s throwing up once I drill down to a single artist – ‘Fans Might Also Like’ is a pretty on-point list of other artists that I might, indeed, like.
Pour one out – Rdio didn’t really do anything that other platforms aren’t doing, but I wanted to call it out as a carefully crafted experience for doing so. Maybe it’s just rose-tinted glasses, but Rdio (specifically desktop Rdio) felt frictionless to me in a way that not many pieces of software achieve.
Rdio also had bios for nearly every artist and reviews for nearly every album, which is not at all important for ‘discovery’, or ‘listening’, but did add an extra dimension to ‘appreciation’. Knowing the narrative of a given artist or album could change the way I perceive the whole thing. It’s a bit of a digital nod to liner notes, or elaborate album packaging, that I appreciate.
8tracks, ThisIsMyJam, Last.fm. All of them seem like they are or were powered almost entirely by user recommendations or user playlists – cool! I can’t deny the power of personal curation in music discovery. Of the three, I’m most interested in ThisIsMyJam in that it made a statement: the primacy of a given track over all others. Not just “I like” but “this is the best”. There’s something super-powerful about that and it’s informed some of my thinking on the future shape of this whole project.
Not noted individually, but all of the platforms in question allow me to explore by ‘newness’ (or more specifically, release date) and also genre. I’ve actually discovered a not-inconsiderable amount of great music just by trawling genre lists on a few different platforms, but that’s reliant on a personal habit of ‘browsing’ which may not exist in all listeners. I think there’s an interesting delta between the concepts of ‘newness’ and release date, in that a track may only have one release date, but could be new to the platform, or new to me, or new in other ways. There’s some deeper slices I’d like to take into ‘new’ as a quality that music possesses.
Pretty much all of the platforms also allow me to create playlists, which I like from a listening standpoint but don’t necessarily have anything to do with discovery, per se. There is definitely an unresolved question about the narrowness or broadness of what 100db might be.
The unacknowledged – or at least unaddressed – problem to overcome with recommendation systems is that (to my mind) it’s hard to distinguish a ‘play’ from a ‘like’. If I play something once or twice I might just be checking it out. If I play it ten times, it’s likely more intentional (counterpoint: over what timeframe?). I’m wondering if there’s a UX model that can capture that kind of meaning from the interaction itself.
Similarly, no platform gives me a Facebook-adjacent ability to indicate any other kind of emotive response to a piece of music – “heart” and “holy shit” are two distinct responses I might want to register. That kind of flexibility can be emergent (the user-level adoption hashtags on Twitter predating platform support being a good example) but I think it’s more interesting as a deliberate mechanism.
Folks on the mailing list had a chance to get an early look at this post, and, indeed, quite a lot of them yelled at me for not including such-and-such platform (thank you) – those’ll definitely show up in a second competitive research post.