ABOUT THE TEAM

Both of us are postgraduate students studying Data and Knowledge Engineering at OVGU Magdeburg. At university, we both have learnt about MIR and deep learning. Richhiey had participated in the AI Song Contest previously whereas Oliver has completed a few projects in MIR in the last semester. We both were always friends and got together for the song contest to explore the possibilities that lie in this space. Our efforts were guided by Prof. Sebastian Stober and Jens Johannsmeier from the AILab at OVGU.

ABOUT THE SONG

This song was based on a story generated by a GPT-2 model trained on pop lyrics. The story was about a girl who has defeated a dragon and is being guided home by the wise one. We really loved the plot and decided to make a song around this idea. With this in mind, we used our AI systems to generate melodies, basslines, percussive elements and drum beats, Most of the elements in the track were generated with tools open sourced by the Google Magenta project. We hope for this track to have a dancy vibe while also showcasing some of the quirks of music generated by computer systems. 

ABOUT THE HUMAN-AI CO-CREATION PROCESS

Our creative goal was to create a track to which we could dance and vibe. To create the overarching story for the entire music we used a GPT-2 model trained on genre related song lyrics. It told us the story of a girl who defeated a dragon and is being guided on a journey home by the wise one. Although GPT-2 wasn’t able to quite completely turn the story into song lyrics, we decided to pick up our favorite bit from it into the song and take the listener on the journey through the music. These picked lyrics were then recorded and resynthesized with several DDSP models to create the AI like vocals in the track - keeping the unique AI generated artifacts 

For the melodies, basslines and rhythms we used tools open sourced by the Google Magenta project. We took this story and decided what genre we’d want to make music in. With the genre as our guide, we generated various musical ideas like basslines, melodies, beats and small sound effects. We also tried using MusicTransformer and PerformanceRNN to experiment with musical ideas. 

We also wanted to include generated environmental sounds in the song generated by SampleRNN models, with which we had varying success depending on the type of environment. The melody that we finally had decided upon had a “beach” feel to it, and unfortunately the generated samples were not good enough to include in the final track. We then decided to use an NSynth Model to crossfade the ocean sounds to the start and end of the song. We found the artifact like sounds generated by this AI model quite interesting and have included it in the track.