Participants 2022
The Everlasters
AI Song Contest 2022 / participantsTEAM / The Everlasters
SONG / Melatonin
TEAM MEMBERS / Luc Leroy, Yann Macé, Emmanuel Deruty, Maarten Grachten
About the TEAM
Luc Leroy (France) and Yann Macé (France)
Luc and Yann have been partners in music production since 1990. Through their company Trak Invaders/Hyper Music they have been producing and mixing tracks for numerous artists (Céline Dion, Jean-Jacques Goldman), films/series (Snatch, The Warrior's Gate), and advertising campaigns (Mercedes, L'Oréal, Nintendo, Adidas). As The Everlasters they release their own creative work, including successful songs like Bang Bang, and Ordinary Tribe.
Emmanuel Deruty (France)
Emmanuel is a researcher in musicology and audio engineering. He is music team manager at Sony Computer Science Laboratories (CSL) Paris, coordinating the team's joint effort to build creative AI tools for music artists.
Maarten Grachten (The Netherlands)
Maarten is a senior researcher and independent machine learning consultant with expertise in music technology and music information retrieval. At Sony CSL Paris he has been responsible for the design and development of several AI-based music production prototypes.
About the SONG
We composed the song "Melatonin" as an experiment with a specific AI tool—BassNet—that generates bass lines in response to existing musical elements. Given the sound and playing modes of the tool, we deemed an urban music style (drill, trap, hip hop) most fitting, creating resolute electronic music with simple, even ecstatic harmonic structures that give BassNet the opportunity to develop striking and clear lines in a conventional production environment. By layering several bass line variations we created richly textured bass parts that feature prominently in the mix.
About the HUMAN-AI PROCESS
We—the artist duo (Hyper Music) and the creators of BassNet (two AI researchers from Sony Computer Science Laboratories, CSL)—arranged a dedicated four-day studio session to work on the song. Since BassNet is designed to respond to musical inputs rather than create bass lines from scratch, Hyper Music prepared some audio material in advance (drums, choir, and keyboard parts). The material was intentionally sparse, to provide space for the bass part.
Hyper Music then created multiple variations using an iterative generation process, where BassNet outputs were conditioned on a mix of the original material, plus BassNet outputs obtained earlier in the process. During the session, CSL made modifications to BassNet in order for the tool to generate more colorful and expressive sound.
Some outputs were processed using other CSL AI sound shaping tools, as well as equalization and effects (pitch shifting, distortion, delay, reverb). Outputs were often layered to create heterophonic textures, and were differentiated to create both low register bass parts, solo parts, and more abstract sonic accents.