This post originally appeared on MIT Technology Review
Old songs, new tricks: Computer-generated music has been a thing for 50 years or more and AIs already have impressive examples of orchestral classical and ambient electronic compositions in their back catalogue. Video games often use computer-generated music in the background, which loops and crescendos on the fly depending on what the player is doing at the time. But it is much easier for a machine to generate something that sounds a bit like Bach than the Beatles. That’s because the mathematical underpinnings of much classical music lends itself to the symbolic representation of music that AI composers often use. Despite being simpler, pop songs are different.
OpenAI trained Jukebox on 1.2 million songs, using the raw audio data itself rather than an abstract representation of pitch, instrument or timing. But this required a neural network that could track so-called dependencies—a repeating melody, say—across the three or four minutes of a typical pop song, something which is hard for an AI to do. To give a sense of the task, Jukebox keeps track of millions of timestamps per song compared to the thousand timestamps that OpenAI’s language generator GPT-2 uses when keeping track of a piece of writing.
Chatbot singalongs: To be honest, it’s not quite there yet. You will notice that the results, while technically impressive, are pretty deep in the uncanny valley. But while we are still a long way from artificial general intelligence (OpenAI’s stated goal) Jukebox shows once again just how good neural networks are getting at imitating humans, blurring the line between what’s real and not. This week, rapper Jay-Z started legal action to remove deepfakes of him singing Billy Joel songs, for example. OpenAI says it plans to conduct research into the implications of AI for intellectual property rights.
