Co.Design

How To Fix Streaming Music Services

Algorithms are a start, but they don't curate, provide context, or persuade you to listen.

Last week I had a familiar experience: I plugged one of my favorite songs (by the excellent drone metal band Earth) into Spotify’s radio feature and their first recommendation was worlds apart from my original. Yes, the genre and vibe were similar, but the song felt completely different to me. Put simply, I hated it. So I skipped ahead to the next song and then the next. Like so many others, I gave up, and instead used Spotify as a way to stream music, not as a means of discovering great new songs.

This is a problem. Online content discovery is a huge opportunity for tech companies, and streaming music platforms are a hothouse of experimentation and dynamism. Unfortunately, the bigger players like Spotify, Pandora, and now Google’s Play Music, have yet to nail the music discovery process. Why is that?

What’s wrong with algorithmic music discovery?

John Paul Titlow's recent article in Fast Company outlines some of the key challenges facing music discovery platforms. The way that many of these services currently attempt to crack the music discovery nut is to break songs down into as many “constitutive” aspects as possible--genre, decade, tempo, etc. It then makes inferences about what a user is looking for next predicated on previous choices.

Google’s new service is a bit more ambitious in that instead of using real people to input these track aspects into their song catalogue, it uses “machine listening.” This means that computers automatically identify the aspects and then feed them directly back into the algorithm. The basic logic of this kind of grouping harkens back to old-fashioned commercial radio stations, where songs were catalogued into categories such as genre (the "Classical" station) or decade (the ‘80s station).

The approach makes sense from a coding perspective--the algorithm must have some way of identifying songs and artists and connecting them to others in their catalogue. However, it has very little to do with how people actually see connections among songs. It may even get in the way of people who are truly interested in music discovery.

How can music platform sites get music discovery right?

Famed innovation expert Everett Rogers grew up on a farm in Iowa. He always wondered why it took his father so long to start using the newly available hybrid seed corn, both drought-resistant and higher producing. He went on to study the question "How do we embrace new ideas?" in academia, and developed the now-famous diffusion of innovations theory. Rogers concluded that behavior change is both a social and a temporal phenomenon.

This is essential for embracing new music: It takes time to commit to a new artist, a new album, just as it takes time to commit to anything new and valuable. According to Rogers, individuals tend to go through five different stages in the process of adopting new ideas. The very first stage is knowledge, or coming into contact with the idea. All music discovery sites get this phase right. They let users "know" about a new song or band.

The second stage is entirely social: Knowledge must be followed by persuasion. Only through persuasion do most individuals make the decision to embrace something new. This decision is then followed by an implementation and a confirmation stage.

What do these stages look like in an actual music discovery process? If an algorithm tells us to listen to a song from Bob Dylan’s Blonde on Blonde, say, we might instantly turn it off because we don’t like his voice. Algorithms have no persuasive power. As much of our most iconic and celebrated music does not have "instant traction"--it is not instantly "likeable"--users on more mainstream music sites are missing out. In fact, many artists specifically design their songs and albums to reveal themselves over time. In a recent segment on National Public Radio, Trent Reznor, former member of Nine Inch Nails, told the interviewer, "I aspire to make a record that sounds better 10 listens in than it does after two--and still, at 50 listens, you're picking out things that add a depth and a thoughtfulness."

If music discovery sites are to succeed, they need to account for the persuasive phase so integral to Rogers’s theory. Great music discovery often requires an impassioned advocate and a commitment of time. (So what if you didn’t like it the first time? I’m telling you, it’ll change your life. Try listening again.)

Let’s take a lesson from the past: Make music social.

Of course, we know all of this intuitively because this is how music discovery worked in the past. Twenty years ago, when people wanted to hear a broad range of new and exciting music, they turned to their favorite DJ. A radio DJ had the advantage of playing a song several times a day, introducing the listeners to the new sounds and allowing them to consider and then embrace an adoption of the music. After the fifth or sixth "play," listeners were able to create their own unique relationship to new music, guided by the trusted curatorial skills of the DJ.

If music platforms have any hope of mainstream success, they have to stop atomizing music into bits and parts, which divorces it of its context. Instead, they need to replicate elements of an essentially social phenomenon, allowing for listeners to embrace new music over time through the persuasions of peer and expert recommendations. Yes, it’s true that a platform like Spotify has shared playlists with a social component but it feels like an afterthought in the program’s design. It’s not an intuitive experience--the interface is clunky and it doesn’t fully capture the aspects that make a personal recommendation so powerful.

Sites that curate or get more subjective get it right.

Fortunately for music lovers, there are several sites out there starting to get the discovery process right:

Turntable.fm, started in 2011, allows users to join chat rooms that stream music from a rotating list of fellow-user DJs. Discussions in the chat room can heat up regarding the musical choices. If enough people decide that a song is “lame,” the app skips to the next song on the playlist.

If Turntable simulates the “real time” social experience of a DJ, Piki, developed by the same designers, provides an alternative to casual streaming. Like Pandora, users simply turn it on and press play. The app will select songs from a playlist curated by friends and trusted influencers. Piki bills itself as the "middle man" between the fully engaged Turntable experience and the completely passive Pandora experience. It is both casual and personal--entirely intuitive--like a radio station designed by your favorite people.

The Echo Nest relies on algorithms for its recommendations but instead of taking the constitutive bits from the music itself--tempo, chord, or genre, for example--it pulls its information from responses to the music. The app mines data from concertgoers, amateur musicians and professional reviewers. This gets the site closer to providing recommendations based on the subjective experience of listening to the music, not the substance of the music itself.

Even better, try to simulate in-store moments and concert-going.

Of course none of these apps have truly mastered the serendipity of a great record store experience. Music lovers know the tactile pleasures of flipping through albums in the bins, making discoveries based purely on the album art or the band name. Mainstream music discovery sites claim they offer this experience as well but their versions are sorely lacking. The album art usually appears on a tiny JPEG somewhere in the midst of a crowded interface and there is no way to do a search based on funny or unusual band names. Sites could better simulate the record store experience by allowing users to enter a virtual space, giving them access to album art in large format and high resolution.

Even more important: I haven’t seen any music sites really bring together the full entertainment experience by offering merchandise, concert tickets, and other relevant contextual information on the bands (tour schedules, Twitter feeds, Wikipedia pages, etc.).

Perhaps all this is a lesson that the art of music discovery can never truly be engineered. Still, we should take inspiration from these, and other, experiments on display. Through them, we can take away larger lessons about user discovery in all product categories. Some day soon, I hope all my consumer sites are bringing me into contact with meaningful products and experiences. Until then, I’ll be the one scrolling through my own library on iTunes for my favorite drone metal bands.

Written by Christian Madsbjerg and Morgan Ramsey-Elliot of Red Associates, a strategy and innovation consultancy based in the human sciences

[Image via Shutterstock]

Add New Comment

23 Comments

  • Will Kreth

    Thanks for this piece, Christian and Morgan -this is exactly the approach our company Critical Metrics took in 2006. http://en.m.wikipedia.org/wiki...

    Bibliometrics and sentiment analysis of global music reviewer opinions may not be as sexy automated as running an algorithm to match songs together, but like the power of word of mouth - there's a human element that code alone can't quite trump.

  • yuval eyal

    discovery is working pretty well. The first sign of it is that most of the music we listen to is new to us, or played by unknown artists. The problem is not discovery. the problem is going back to a track you bumped into while listening to a curated stream  (manually or automatically), and saving it for later. as music lovers use various music services such as the ones mentioned in the post and in comments, they need a tool to collect tracks that touched them and add them to their music collection. just like downloading a track in the 00's or buying a CD in the 90's. in a streaming world, download means nothing. 

    Sharing is another problem. Spotify premium users can't share a track with Rdio premium user, although both are paying for the same track.
    We've created whyd.com , a service that let's you collect and share tracks you love from various streaming services. our community of users build their own music collections, but can also access other people's collections based on social connections or musical taste.
    discovery on whyd is based on human curators, but i think that algorithmic discovery can be as good a human in some cases. sometimes i want to listen to someone i trust with a musical taste, and sometimes i just want to go with a genre i'm in the mood for. everything works.
    here is an invite for whyd:  http://whyd.com/invite/50b5f0b...
    (we are still on invite only)

  • Miguel Banuelos

    One of the main issues that services have in terms of discovery is that they are trying to be everything to all people. The reason record stores and labels have always been fractured along genre as well as 'scene' lines is that each of these groups (and every person) discovers music differently and need different reasons to listen to something new. Spotify is not offering things to a person who already knows that a band like Earth even exists. That's not for you. Just like if you went to a Best Buy and asked for bands that sound like Earth, they would probably have no idea what you required either. Perhaps if these streaming services broke things down in proper ways to serve proper kinds of users, they could give you what you want. But that isn't algorithmic, it's human.

  • Ricardo Silva

    On account of this : "Even more important: I haven’t seen any music sites really bring together the full entertainment experience by offering merchandise, concert tickets, and other relevant contextual information on the bands (tour schedules, Twitter feeds, Wikipedia pages, etc.)."

    I want to suggest musikki.com , they have been around for a while, but just released a new beta version recently. This is just the beginning, but it seems promising really.Also I really like their feature of "social player" .. it picks music that was shared on your facebook by some friends or pages. I use it to create playlists from friends who have I know have a great taste in music, or from pages that curates music.

  • Matt Fiedler

    I'm am the co-founder of a serivce called Vinyl Me, Please (http://vinylmeplease.com) and this is exactly what we're trying to accopmlish.  We hand pick a new record each month and try to create an experience around getting new music that is equivalent to that of receiving a birthday present in the mail every month.  We work very hard to give a personalized touch on music discovery with both analog and digital formats in mind.

  • Michael Writhe

    Excellent insights... And right along the lines of what we are trying to do over at Grooveshark with the broadcasts feature. I am just one of hundreds of DJ's curating music constantly for our listeners. Stop by and check us out sometime... I promise we won't let you down :)

  • conbdebolio

    So... Rdio didn't even get a nod? I don't know how good/bad they are from the discovery standpoint, but the interface is BY FAR the best one. It FEELS like music.

  • Trishanth

    The Last.fm reccomendation service is actually quite good. I use an app that uses the Last.fm API called Lazify and I have discovered relevant music based on what I already listen to. I agree there is work to be done, but its not a disaster.

  • Skye Pathare

    Try Deezer - it's excellent. They employ local teams of music editors to compile playlists and make recommendations and don't use algorithms, and so far they've been pretty accurate and helped me to discover heaps of new bands

  • Graham

    This article should really be called "One guy's singular disappointing experience with Spotify's radio feature". 

    Come on, no mention of Songza or 8tracks? Both of which are based on user-curated playlists *specifically* designed to deliver a specific mood, mix, vibe, etc. 8tracks alone has over 20,000 mixes, from which I've discovered dozens of awesome new bands at parties, in the car with friends, or all by my lonesome on a long commute home.  

    Also, Pandora's Music Genome Project is pretty damn well done, all things considered. It involves the manual processing by professional musicians of 400 something tags per song (Wikipedia that shit). Maybe you feel it still doesn't replace your local indie record clerk, but it works amazingly well for me. Also, what about Grooveshark and Last.fm, both of which offer their own radio streaming algorithms? Last.fm is the grandaddy of them all, and its song-matching process -- algorithm or not -- really knows what it's doing. Unlike your claim that you haven't seen a music site "really bring together the full entertainment experience", Last.fm also offers a photo slideshow, tour info, concert tickets, and wiki information on every band, just as you describe."Perhaps all this is a lesson that the art of music discovery can never truly be engineered." -- Uh, yeah, no. More like, perhaps this is a lesson in Fast Co's transparent formula for pumping out articles based one guy's arbitrary opinions wrapped around a bogus thesis ("music discovery must be social to work!"), with a clickbait headline and shameless plugs for obscure services that probably paid for the mention.

  • Graham

    Yeah, so thanks for censoring my comment. Nice to know Fast Co can't take pointed criticism either. What a farce... 

  • Jules Lombard

    There is already a streaming music service out there that are not curated by some algorithm or machine language or what not, and has been around longer than Pandora or Spotify ... SLACKER RADIO.  They're available in both iOS, Android, the Web, Windows 7, Roku, etc ....

    Give it a try and you guys be the judge ...

  • Jesse Cotterman

    I would say Spotify suffers from "lack of discovery" much more than the other apps you mention.  Spotify is pretty much worthless unless you want to listen to a specific artist, song, or album.
    On the other hand, I feel Pandora is pretty good at discovering music - especially if you periodically add stations based on newly discovered songs or artists that come up - it grows the selection organically, with minimal "curating" by the listener.

    The persuasion factor has little benefit for me personally - if I don't like a song, I don't like a song - whether it's voice, tempo, or whatever.  If a DJ plays a song I don't like 3 times a day, it doesn't make me like it - it usually makes me dislike it even more!

    The "serendipity of a great record store experience", IMHO also often sucks.  You pick something based on cover art or album title, rather than on the music itself, and odds are about 3:1 that it won't be that great.  Occasionally, you get lucky, and it is great, and that's awesome - you feel like you just won the lottery - but just like the lottery, you spend a lot of moneyto win.  Why? Because you are "judging the book by its cover" - not selecting the music based on the music.  I've found the best way to discover is with sites like Pandora.  The best way to validate before purchasing is to listen to the rest of the album on a site like Spotify.  Then, go make your happy purchase - or don't, and just enjoy this world of great and practically free music we live in today...

  • Graham

    This article should really be called "One guy's singular disappointing experience with Spotify's radio feature". 

    Come on, no mention of Songza or 8tracks? Both of which are based on user-curated playlists *specifically* designed to deliver a specific mood, mix, vibe, etc. 8tracks alone has over 20,000 mixes, from which I've discovered dozens of awesome new bands at parties, in the car with friends, or all by my lonesome on a long commute home.  

    Also, Pandora's Music Genome Project is pretty damn well done, all things considered. It involves the manual processing by professional musicians of 400-something tags per song (Wikipedia that shit). Maybe you feel it still doesn't replace your local indie record clerk, but it works amazingly well for me. Also, what about Grooveshark and Last.fm, both of which offer their own radio streaming algorithms? Last.fm is the grandaddy of them all, and its song-matching process -- algorithm or not -- really knows what it's doing. Unlike your claim that you haven't seen a music site "really bring together the full entertainment experience", Last.fm also offers a photo slideshow, tour info, concert tickets, and wiki information on every band, just as you describe."Perhaps all this is a lesson that the art of music discovery can never truly be engineered." -- Uh, yeah, no. More like, perhaps this is a lesson in Fast Co's transparent formula for pumping out articles based one guy's arbitrary opinions wrapped around a bogus thesis ("music discovery must be social to work!"), with a clickbait headline and shameless plugs for obscure services that probably paid for the mention. 

    Nice try, but fail. 

  • Postcyberpunk

    For non-pop music:

    Soundcloud effectively allows for the organic curation of music with their stream feature.  I think they are on the right track.  I can see what other music fans I follow are 'reposting'. 
    Mixcloud does a good job by facilitating the upload of DJ sets.
    Rdio offers the ability to stream a record label's discography, I find labels often have a lot of familiar sounding music that is all at a similar production level.