3 ms·

Infinite music is interesting from the angle that the music that we value is connected to our cultural and social experience. How can we cherish a song that has

Infinite music is interesting from the angle that the music that we value is connected to our cultural and social experience. How can we cherish a song that has never been heard before and will never be heard again, which means it is deprived of social context that would give it meaning? One answer would be to create music that shares its roots with music that the listener already knows. This music could be enjoyable, but you can't exactly sing along to a melody you're hearing for the first and last time, so it has more limited engagement potential. This is an approach to composition that you learn when you study chord progressions and other elements in music theory, and it's what I'm sensing when I listen to the MusicGen outputs. To draw from greater cultural context, you can incorporate folk and popular melodies that are widely known. Musicians love this trick. "Immature artists copy, great artists steal." MusicGen seems capable of doing this, too. To promote a novel melody as something that listeners deeply cherish, or to innovate at the level of the theory, the social context has to be built up around the content after it's generated. E.g., when introducing a new song on the radio, a common trick is to play it between songs that are already popular; building up co-occurrences with songs that already have cultural significance. My challenge to Meta would be: can you use your platform to transform some of the model's novel outputs into familiar popular music? It would be an important cultural milestone if an AI-generated melody became a familiar tune that would be played in the café, recognized, and enjoyed.

visarga 3y ago
> It would be an important cultural milestone if an AI-generated melody became a familiar tune that would be played in the café, recognized, and enjoyed New generative music benchmark - popularity.
- odyssey7 3y ago
  I can see why some might think this is silly, it seems unscientific. But what exactly is human-level performance in music, and how can we detect that? If music is an artifact appreciated by listeners, then any metric apart from whether the music is listened to would be a proxy — though I can appreciate the perspective of creating for the artist’s own sake, without a need to share the creations. Popularizing some of the model’s outputs would reveal their merit against human-produced music, by allowing them to succeed or fail in attaining the same quality of cultural significance. Then again, if the artist here is the AI research team and the audience is AI enthusiasts, then the music has succeeded in being heard and attaining cultural significance. It has been remarked that Schoenberg’s music was more often defended than listened to — music written by a theorist for an audience of theorists. I am a true fan of Schoenberg’s, though I can hear that the example outputs of this work are music that is meant to be accessible to the everyday listener.
pcthrowaway 3y ago
I've thought about this same idea before, but especially as it relates to film/television, and writing. Think of your favourite TV show. If when you first watched that show, you were told no one else had ever watched it, or ever would watch it, would your engagement with it be the same? Part of our enjoyment of art is the shared cultural context. Maybe saying "art" here isn't the salient thing. Maybe it's our engagement with ideas. I personally haven't even considered this concept as much in relation to music, because while I do love to deeply engage with music and the shared narrative behind it, both real and imagined, I also just like to put on music that sits in the background as a tool to drown out noise while I'm working, walking, etc.

visarga 3y ago

> It would be an important cultural milestone if an AI-generated melody became a familiar tune that would be played in the café, recognized, and enjoyed New generative music benchmark - popularity.

odyssey7 3y ago
I can see why some might think this is silly, it seems unscientific. But what exactly is human-level performance in music, and how can we detect that? If music is an artifact appreciated by listeners, then any metric apart from whether the music is listened to would be a proxy — though I can appreciate the perspective of creating for the artist’s own sake, without a need to share the creations. Popularizing some of the model’s outputs would reveal their merit against human-produced music, by allowing them to succeed or fail in attaining the same quality of cultural significance. Then again, if the artist here is the AI research team and the audience is AI enthusiasts, then the music has succeeded in being heard and attaining cultural significance. It has been remarked that Schoenberg’s music was more often defended than listened to — music written by a theorist for an audience of theorists. I am a true fan of Schoenberg’s, though I can hear that the example outputs of this work are music that is meant to be accessible to the everyday listener.

pcthrowaway 3y ago

I've thought about this same idea before, but especially as it relates to film/television, and writing. Think of your favourite TV show. If when you first watched that show, you were told no one else had ever watched it, or ever would watch it, would your engagement with it be the same? Part of our enjoyment of art is the shared cultural context. Maybe saying "art" here isn't the salient thing. Maybe it's our engagement with ideas. I personally haven't even considered this concept as much in relation to music, because while I do love to deeply engage with music and the shared narrative behind it, both real and imagined, I also just like to put on music that sits in the background as a tool to drown out noise while I'm working, walking, etc.