OpenAI today introduced ‘Jukebox,’ an open-source AI system for generating novel music complete with lyrics and vocals. Simply put, it’s the most impressive music-writing AI we’ve seen yet.
Jukebox was trained on massive datasets of music across nearly every genre so that it can create completely fresh songs that, in many cases, sound very much like the artist(s) it was trained on.
Much like other networks learn to transfer the style of a painting or other artwork from one image to another, Jukebox figures out how to imitate the style and genre of music. It also attempts to recreate the lyrical and singing style of specific singers.
Per the OpenAI team’s research paper:
We show that our models can produce songs from highly diverse genres of music like rock, hip-hop, and jazz. They can capture melody, rhythm, long-range composition, and timbres for a wide variety of instruments, as well as the styles and voices of singers to be produced with the music.
Rather than describe it any further, you’ll just have to hear it for yourself:
OpenAI‘s been working on generating music for a few years now. Its previous effort, MuseNet, effectively produced song-length MIDI tracks (basic tones and compositions), but as far as we know there’s nothing out there that can create songs with full vocal parts in nearly any genre like Jukebox.
The applications for Jukebox are mostly academic right now. Unless there’s a category for “odd, machine-generated pseudo-songs,” we won’t be hearing any of Jukebox’s tracks on the charts anytime soon. It lacks the prowess to follow basic music structure. You’ll notice there aren’t any repeating hooks or lyrical bridges.
But, we’ve seen this kind of roll out from OpenAI before. They come up with a somewhat interesting use-case and then, over time, they beef up the models by increasing the parameters. Right now, Jukebox does an impressive imitation of someone mumbling through a fairly decent artist impersonation (seriously, the AI’s Elvis and Sinatra are spot-on). But perhaps future iterations will bring brand-new tracks that are indistinguishable from artists’ actual work.
Unfortunately, you can’t get Jukebox to generate new music just yet. It takes about nine hours to render a single minute of music with it currently, which puts any large-scale interactive uses out of reach for the time being.
Luckily though, OpenAI has gone through the trouble of creating a browser you can use to sift through the dozens and dozens of song samples featuring music in the style of everyone from Michael Jackson and Celine Dion to Aesop Rock and NoFX (shout out to the OpenAI dev with rad taste in music).
Get the Neural newsletter
Greetings Humanoids! Did you know we have a newsletter all about AI? You can subscribe to it right here.Follow @neural