Belgium - Polaris

This team consists of an experienced data scientist, a professional musician who played in bands and two physicists -- one who also plays drums. They all work for the same digital product studio.

In this Belgian entry, Polaris composed their song with Google's open-source Magenta, trained  using 1.5 million unique online MIDI files. The nice thing about this is that anyone can play around with it, since it’s open source. The melody, rhythm and bassline were simultaneously generated in one go by the AI.

The team used data provided by the AI Song Contest organization to generate new lyrics from the 200 existing ones. For this, they used the GPT-2 neural network from OpenAI. Matching the lyrics with the melody turned out to be difficult to automate, so the team did that themselves.

The singer’s melody is generated by the AI, but they used a human to perform the song. The vocalist is Maarten Van den Bossche, and Sebastiaan van den Branden plays the AI-generated bassline.The rest is performed by synths and drum computers.

In other words, this is a collaboration between man and machine. The lyrics, melody, bass parts and rhythm are all computer generated, but the song is put together and partly performed by humans.

‘We had hoped to lean completely on AI,' the team writes. ‘It would be great to conjure up a complete, rocking song at the push of a button, but unfortunately that's not yet possible. That said, we’ve found that AI and humans together make a great team.’