Every year, participants from primarily European countries put on an absolute spectacle for the world in the Eurovision Song Contest. Fireworks, glitter, costume changes, key changes, wind machines, fake rain… you name it, it has probably happened in the last 64 years of the competition. However, due to the Covid-19 pandemic, this year’s action, scheduled for May 16, will not go ahead in The Netherlands, last year’s winning country.
But here to fill this void is the AI Song Contest, organized by Dutch public broadcaster VPRO. For the first time, 13 teams from Europe and Australia have harnessed the creative power of artificial intelligence to generate the next Eurovision hit. And just like the true contest, you (the public) can help to decide who gets douze points and who walks away with nil points.
Computer scientists have paired with rappers, machine learning experts with folk musicians to train AI systems to compose an instant 3-minute classic. Most teams began by loading their systems up with the repository of Eurovision songs and then used various open-source tools, such as Google’s Magenta project, to help generate new lyrics and melodies.
“In our case, it looks at all the music that ever came from Eurovision,” Ashley Burgoyne, a member of one of the Dutch teams, explained in a YouTube video. “We tell the computer: listen, listen again, and listen some more, until it fully understands all the patterns and all that music. That way it will learn eventually how to write new music itself.”
More often than not, the AI systems came up with some left-field ideas. The lyrics generator built by Dutch team Can AI Kick It created a whole new word “Abbus,” which it defined as “a nascent cloud,” whereas German Team Dadabots x Portrait XO found that their system produced fairly topical lyrics about infections and diseases, such as “my species targets vital signs, path of contagion, bodily fluids.” Naturally, the team incorporated these into a heavy-metal love song between a human girl and AI boy during a severe epidemic.
As for the performance of these masterpieces, some groups chose to have human voices blended with robotic sounds, whilst others fully committed to the good, the bad, and the ugly of AI with expressionless synthesized voices. The Australian entry, Uncanny Valley, even chose to train their neural network with audio samples of koalas, kookaburras, and Tasmanian devils to mark the Australian bushfires that had devastating consequences for its inhabitants.
Voters are asked to rate each AI song in four categories: song, lyrics, originality, and Eurovisioness. The deadline to do so is May 10 and these scores will be combined with those from a panel of experts, who are also looking out for elements like the use of data, song structure, and the extent to which humans have been involved.
Whilst the prize may only be a slot on Dutch national radio, the whole project aims to explore the creative possibilities (and limitations) of AI. Judging by the entries, next year’s Eurovision contestants may well include some computer scientists.
[H/T: Science Magazine]
