ABOUT THE TEAM

Comprising Leila, LJ and Dav; an artist, musician and technologist joining forces to explore how much AI can inspire creativity. LJ, pianist and inventor, first learned AI Composing to take her mind off chemo treatments which had affected her ability to use her hands and mind. After successfully ‘mashing up’ two artists at once, she was hooked - fascinated by how AI made different musical decisions based on its training. Leila – a creativity consultant and former digital artist specialising in AI theory – created powerful AI-generated imagery based on the lyrics. Dav, a technology strategist exploring the application of AI, ex- IBM Services Research Director, interested in exploring AI’s potential in creative approaches to problem solving. All three are intrigued by the process of making music. Together the team found ways to incorporate machine learning into the creative process while aiming for the final song to have wide musical appeal.

ABOUT THE SONG

Originally the 3-person Smorgasborg Collective intended to write a simple dance track but GPT-3, the text generating AI, had other ideas. One prompt resulted unexpectedly in a beautiful many-layered pop song which the team feel is about love and support - set in space. We’re hoping for a true mix between human-inspired / machine generated and machine inspired / human-generated sound, feeding many ideas and outputs back into the process where possible. We used Machine Learning synths and AI-created imaginary soundscapes generated from images. We also used AI to solve where the melody should go at the end of the chorus. We created Imagery using Deep Dream where the seed picture was also inspired by the AI lyrics - the artwork inspiring more orchestration - you get the picture. Mainly, we enjoyed making human friendly music with help from machines and hope you enjoy listening to it too.

ABOUT THE HUMAN-AI CO-CREATION PROCESS

We generated lyrics from different prompts into GPT-3’s AI engine. Many responses - or completions - weren’t suitable for songwriting. Some were too racy for radio, others too simplistic, some didn’t feel like songs - but one stood out - the lyrics  are the unchanged original output.  

The melody came easily during a video call - but one line refused to sing! We entered the troublesome lyrics into Jukebox  with an AI trained on Adele’s songs in case it could solve the issue. Aier many akempts leaving it to create on its own, we  chose co-composing, checking on the model every 4 seconds and choosing a favourite option each time. This unfolded  different musical choices, greatly augmenting the tune and fixing the problem!  

The orchestration contains layered synths trained on over 300,000 sounds - over 1,000 instruments. For the chorus, we  generated beautiful imaginary firework sounds for the background by introducing an image of space to an AI expecting an earth landscape. Deep dream soiware informed the artwork, pumng space pictures inspired by the song through a neural network and exploring the results. Those with a producer ear might hear samples of old technology like fax machines  spread binaurally through the mix - inspired by the ‘number stations’ broadcasting almost-understandable information, like  an AI in training. But we simply wanted to build an emotional song capable of generating a human response with a wide  enough musical appeal for radio play, for it not to be immediately apparent that AI was a writing partner.