Reviving Viva

Artificial Intelligence, AI-Driven Music Production

Personal Project

Tool : Jukebox, Adobe Audition, Aiva, Splitter AI, DallE2

Timeline : 3 weeks (June 2023)

I have created an AI-generated music of Coldplay's "Viva La Vida" using Jukebox, Adobe Audition, and Splitter AI. Due to the limitations of generating music with Jukebox for long audio files, I divided the music into four sections: intro, chorus, bridge, and outro, creating separate audio files of around one minute each. These files were then used as training input for Jukebox. However, every time I trained the model, the generated music had different styles. So, I went through multiple iterations and used Adobe Audition to edit the music and make it flow smoothly. I meticulously edited the music at the millisecond level to minimize any noticeable transitions and applied noise reduction techniques to remove noise from the generated music, giving it a vintage audio feel. Additionally, for automatically generated lyrics that were not understandable, I used Splitter AI to remove lyrics as much as possible from the music.

The cover is generated by DallE AI Image Generate Tool.

📍 Process

The structure of Coldplay's "Viva La Vida" has a repetitive nature with a consistent rhythm, which made it suitable for generating high-quality music using Jukebox. The learning model was able to implement repetition quite effectively, which was a successful aspect of the process. However, since the quality of the generated music varied significantly with each iteration, there was no guarantee of consistently achieving successful iterations. It was observed that generating more than three samples at once resulted in significantly lower quality, so it was preferable to generate 1-2 iterations at a time.

Furthermore, the music files contained a considerable amount of noise, resulting in poor audio quality. Removing the noise and seamlessly connecting each section without noticeable transitions required a significant amount of time and effort.