This is what happened when AI tried to write a Joe Bonamassa-style blues track
The Jukebox neural net can generate songs completely from scratch in the style of a chosen artist - the results might surprise you
The world of artificial intelligence is developing at an unprecedented rate. Once the realm of science fiction and fantasy, AI-powered tools are now accessible to all of us, including music-makers.
One of the most interesting questions raised by this rapid acceleration is whether AI can ever generate music that's indistinguishable from the real thing. One of the closest attempts we've heard to date has come from the team at AI research company OpenAI, which has produced a neural net that generates music from scratch as part of its Jukebox project.
In contrast with AI music generators we've heard previously, that emulate the style of famous composers by generating music in the form of MIDI to be played by existing software instruments, Jukebox takes things one step further by generating raw audio files.
OpenAI trained its neural net on 1.2 million songs, providing the system with audio, lyrics and metadata such as artist name, genre and year of release. Once the model learned this data and was then provided with lyrics previously generated by an AI language model, the researchers were able to ask it to generate music in a particular genre and in the style of a chosen artist, with surprisingly accurate results.
This leads us to Jukebox's AI-generated Joe Bonamassa-style blues rock, which you can hear below. Not bad for a computer program, don't you think? If we'd heard it through a car radio, we might have just fallen for it...
There's a whole library of examples to browse on the OpenAI website, including AI-generated music in the style of Elton John, Britney Spears, Rick Ross and AC/DC. Our personal favourite is a surreal Katy Perry-style ballad that sounds like it's being sung in Simlish.
Due to the limitations of the technology, the tracks sound a little like they're coming out of a cheap and battered Bluetooth speaker. Interestingly, although the model has managed to quite accurately mimic genres and styles, generate conventional chord structures and even solos, it struggled to produce "familiar musical structures" such as repeated choruses.
Get the MusicRadar Newsletter
Want all the hottest music and gear news, reviews, deals, features and more, direct to your inbox? Sign up here.
While it doesn't look like Jukebox will be replacing musicians anytime soon, the project is an impressive indication of the ever-growing power and sophistication of artificial intelligence.
I'm MusicRadar's Tech Editor, working across everything from product news and gear-focused features to artist interviews and tech tutorials. I love electronic music and I'm perpetually fascinated by the tools we use to make it. When I'm not behind my laptop keyboard, you'll probably find me behind a MIDI keyboard, carefully crafting the beginnings of another project that I'll ultimately abandon to the creative graveyard that is my overstuffed hard drive.