Participants 2024
Asian Culture Research Team
AI Song Contest 2024 / participantsTEAM / Asian Culture Research Team
SONG / Overfitting
About the TEAM
Four guys from China, three of them are doing music AI in industry, one of them is doing Fashion AI research in academia.
About the SONG
The theme is overfitting. Modern civilization has provided us with choices, but also restricted our choices. In order to meet the expectations of society, we are all going through the process of overfitting. However, this process itself is a process of institutionalization, and many people have also become alienated as a result. We want to break through the boundaries by using chaotic noise to create music. Over time, the songs gradually incorporate elements of noise, and the style also gradually becomes more frenzied. To achieve this, we use a large amount of random components in our creative logic. At the same time, we also introduce a large amount of randomness as our prompts in the loop generation process. We also keep the song's chords relatively simple, which also implies a line of overfitting in everyone's growth experience.
About the HUMAN-AI PROCESS
1.The majority of the audio sounds, including arp, bass, breakbeat, drum loops, FX, pad, piano, random drums, synth, and techno synth, are generated using the Stable Audio tools (https://github.com/Stability-AI/stable-audio-tools).
2.The lyrics are generated by a combination of Suno and ChatGPT-4.
3.The melody and chords are generated using "Symbolic music generation conditioned on continuous-valued emotions" and "MUSIC FADERNETS: CONTROLLABLE MUSIC GENERATION BASED ON HIGH-LEVEL FEATURES VIA LOW-LEVEL FEATURE MODELLING", with some samples taken and further refined.
4.Some of the drums are generated using the DeepDrummer tool (https://github.com/mila-iqia/DeepDrummer), which can adapt to the user's preferences in real-time. The authors have made some modifications to the DeepDrummer generation logic to introduce more randomness, including the addition of random FX.
5.The vocal effects are created using NeuCoSVC (https://github.com/thuhcsi/NeuCoSVC), which performs voice conversion based on the main vocal, with the addition of delay, reverb, and pitch-shifting effects.
6.The album cover is generated using Stable Diffusion 3-2b, with the prompt "album cover, 4 Humanoid sculpture with black paint texture, wherein 3 non-binary and 1 male, surface of sculpture is Metal reflection, pose of sculpture like the cover of kraftwerk, clean white background, Best Quality, 8K,HD".