> 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
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.
New generative music benchmark - popularity.