Participants 2022
Cactus
AI Song Contest 2022 / participantsTEAM / Cactus
SONG / Not a symphony
TEAM MEMBERS / Luzie Ahrens, Zeyu Yang
About the TEAM
We are team Cactus: Luzie from Germany and Zeyu from China. Both of us started the masters program in Audio communication and -technology at TU Berlin around one years ago. Both of us have some ML background but never used it in combination with music. We saw this contest as an opportunity to try out ideas and explore the possibilities of AI tools in a creative way.
About the SONG
We wanted to make the songwriting process as easy as taking care of a cactus. Our goal was to automatically create ideas as a starting point for musical pieces. The result is a material bank containing MIDI clips and audio files. The MIDI clips offered a lot of room for human creativity, whereas the generated audio files revealed more expressiveness from the side of the AI. In our case, the AI has a pretty dystopian, alien-like sound to itself and added a dark emotionality to the song. If we would create an entity based on the sound, this AI would certainly not be easy to become friends with. If anything, it would be a wild creature living in a dark forest far from civilization, spotted only once every hundred years. It would be mentioned in children's books, along with a warning to never go to the forest alone.
About the HUMAN-AI PROCESS
We saw AI as a creative collaborator rather than a competitor. The songwriting process was a constant exchange of ideas between AI and us. We started with an audio recording of a piano piece from Shostakovich (Prelude 24, op. 34) and used different AI tools to manipulate it. We began by using a Piano Transcription Tool to automatically generate a MIDI file from the audio recording. Furthermore, we generated more MIDI clips from the transcription with Music Transformer, turned the generated files to the audio again and did some more training with RAVE. Some generated MIDI clips were loaded into MIDI-DDSP to produce violin solos. We also used a pre-trained model from RAVE, which is capable of transforming audio into vocal-like sound. This is what ended up sounding like an aggressive alien.
After generating enough material for our sound bank, we chose MIDI clips and audio files that we liked and tried to showcase all of our experimental ideas in the song. With the help of all the processes mentioned above, we turned a masterpiece of classical music into an entirely new form. And it is most certainly "not a symphony".
AI also helped us to generate the cover picture for the song. We used Disco Diffusion to create images based on the keywords "Shostakovich playing synthesizers in the galaxy" and "Shostakovich conducts a symphony orchestra of cacti". Once again, AI came in as a helpful hand when our simple minds ran out of creative energy.