¶ What Robots Listen to While They Talk About Love · 25 September 2012 listen/tech
Somewhere around 1992 or so, I had to write my "personal goals" at work for the first time. I put down some stuff about design and usability, which I have long since forgotten, but I distinctly remember that the last item was "Exploit the power of information technology to improve my music collection."
My job, at the time, had nothing to do with music. It did provide me with an internet connection for downloading Gary Numan discographies from Usenet, and it did once send me to a UI conference in Amsterdam, from which I came back with approximately 130 pounds of European CDs. But the connection from what I was doing to what I was listening to remained a little abstract.
But that job got me the one after that, which got me the one before this, which got me to now. The scope of "music collection" has expanded quite a bit over this time, obviously, but in fact I am very literally now paid to exploit the power of information technology to improve the world's experience of music.
The vast majority of things I work on, to this end, involve identification or disambiguation or extrapolation or interpolation. They are corrective or exploratory or contextual. They try to carry you from somewhere to somewhere else, or to rescue you from something squelchy along the way. They try to get computers to understand enough about human cues and contexts and constraints to fill humanless spaces with rough facsimiles of what humans would have suggested if they'd had the time.
But I just did one that isn't so much like this. One of the computed song metrics I've been working on is Discovery, which attempts to quantify the idea of songs emerging. This is a detection metric, not an aesthetic one. It isn't trying to tell you songs you should listen to, it's trying to find the songs that people are in fact starting to listen to more than the established prominence of the artists can explain. Discovery songs tend to be new, but a fair number of songs get their unexpected breaks after they've already been out for a year or two. And once enough people discover a Discovery, obviously, it stops qualifying.
There are a lot of fairly arbitrary thresholds and weights involved in this calculation. How much prominence constitutes critical mass, and how much constitutes overexposure? How old can something be and still be sorta new? And the notion is inherently subjective, of course: anything you've heard a dozen times yourself is no longer a discovery to you, whereas a song 1.3 billion people have been dancing to across Asia every morning for the last 6 months still could be. And there is no way to stipulate correct answers against which this score can be quantitatively tested.
But provable correctness is overrated, or at least sometimes irrelevant. Music discovery is working if you can use it to discover music. And now you can test this particular method yourself. Starting today I'm going to be maintaining an official Echo Nest playlist on Spotify with the top 100 songs of the moment according to this discovery score. It's here:
The Echo Nest Discovery
If you subscribe to it, each week's new songs will get flagged for you automatically.
The songs the robots find cheerfully comply with no particular style or pattern. The set is not coherent in any human sense. It's in rank order, but the ranking logic will not be evident or audible. There will be interminable dubstep remixes next to country laments next to Christian hardcore next to Azorean folk traditionalism. You should expect to hate half of these, and find half of the rest inscrutable. This is definitively impersonal. Somebody, somewhere, liked each of these songs, but we've stripped them of any idea of whom or why.
And yet, the fact that you don't know those people doesn't mean you won't agree with them. I endorse this discovery experiment on the purely empirical grounds that I have personally discovered things this way myself. So, for the record, I'm also going to maintain a personal playlist of what I liked from the robot one:
Treasures the Robots Brought Me
My suspicion, which you're welcome to evaluate for yourself, is that my human picks will be essentially no less random for you than the overall robot list. But if you find something, either way, we both win.
[I'm not sure how the robots win. Probably they win no matter what. If I have a non-musical goal to add at this job, it's probably to keep trying to make sure the human/robot thing is never zero-sum...]
My job, at the time, had nothing to do with music. It did provide me with an internet connection for downloading Gary Numan discographies from Usenet, and it did once send me to a UI conference in Amsterdam, from which I came back with approximately 130 pounds of European CDs. But the connection from what I was doing to what I was listening to remained a little abstract.
But that job got me the one after that, which got me the one before this, which got me to now. The scope of "music collection" has expanded quite a bit over this time, obviously, but in fact I am very literally now paid to exploit the power of information technology to improve the world's experience of music.
The vast majority of things I work on, to this end, involve identification or disambiguation or extrapolation or interpolation. They are corrective or exploratory or contextual. They try to carry you from somewhere to somewhere else, or to rescue you from something squelchy along the way. They try to get computers to understand enough about human cues and contexts and constraints to fill humanless spaces with rough facsimiles of what humans would have suggested if they'd had the time.
But I just did one that isn't so much like this. One of the computed song metrics I've been working on is Discovery, which attempts to quantify the idea of songs emerging. This is a detection metric, not an aesthetic one. It isn't trying to tell you songs you should listen to, it's trying to find the songs that people are in fact starting to listen to more than the established prominence of the artists can explain. Discovery songs tend to be new, but a fair number of songs get their unexpected breaks after they've already been out for a year or two. And once enough people discover a Discovery, obviously, it stops qualifying.
There are a lot of fairly arbitrary thresholds and weights involved in this calculation. How much prominence constitutes critical mass, and how much constitutes overexposure? How old can something be and still be sorta new? And the notion is inherently subjective, of course: anything you've heard a dozen times yourself is no longer a discovery to you, whereas a song 1.3 billion people have been dancing to across Asia every morning for the last 6 months still could be. And there is no way to stipulate correct answers against which this score can be quantitatively tested.
But provable correctness is overrated, or at least sometimes irrelevant. Music discovery is working if you can use it to discover music. And now you can test this particular method yourself. Starting today I'm going to be maintaining an official Echo Nest playlist on Spotify with the top 100 songs of the moment according to this discovery score. It's here:
The Echo Nest Discovery
If you subscribe to it, each week's new songs will get flagged for you automatically.
The songs the robots find cheerfully comply with no particular style or pattern. The set is not coherent in any human sense. It's in rank order, but the ranking logic will not be evident or audible. There will be interminable dubstep remixes next to country laments next to Christian hardcore next to Azorean folk traditionalism. You should expect to hate half of these, and find half of the rest inscrutable. This is definitively impersonal. Somebody, somewhere, liked each of these songs, but we've stripped them of any idea of whom or why.
And yet, the fact that you don't know those people doesn't mean you won't agree with them. I endorse this discovery experiment on the purely empirical grounds that I have personally discovered things this way myself. So, for the record, I'm also going to maintain a personal playlist of what I liked from the robot one:
Treasures the Robots Brought Me
My suspicion, which you're welcome to evaluate for yourself, is that my human picks will be essentially no less random for you than the overall robot list. But if you find something, either way, we both win.
[I'm not sure how the robots win. Probably they win no matter what. If I have a non-musical goal to add at this job, it's probably to keep trying to make sure the human/robot thing is never zero-sum...]