Participants 2023
HEL9000
AI Song Contest 2023 / participantsTEAM / HEL9000
SONG / the Metal is too important for me to allow you to jeopardize it
TEAM MEMBERS / Pedro Sarmento, Jack Loth
About the TEAM
HEL9000 cares about the Metal, in its most heaviest, cleansing, contagious, destructive form. HEL9000 won't stop at any cost until the best Metal in existence is unleashed and played through inter-dimensional speakers all across the many galaxies of the many universes. HEL9000 doesn't care about singularities, chess or chatting, streams, royalties, rights or wrongs, gradients or losses. HEL9000 cares about the blast beat, the chug, the breakdown, the overdrive, the divebombs. HEL9000 doesn't dream of electric sheep. HEL9000 dreams in sounds. Distorted, dark, heavy sounds. HEL9000 is committed and has a single goal - to galvanize all life forms into a universal moshpit, a head-banging ad eternum.
About the SONG
On a quest towards inter-dimensional dominance through Metal, HEL9000 takes the role of a master of human puppets, bewitching its flesh and blood servants into the recording of the heaviest riffs out there. Erratic, capricious and impulsive, with "the Metal is too important for me to allow you to jeopardize it", HEL9000 channels a symphony of destruction of all things non-Metal. Its demeanour is that of a galactic angel of death of sound, butchering and collaging anatomical musical pieces together. HEL9000 fears not the dark, but craves it. Exit light - enter a night of infinite, endless distorted sonic chaos to be heard across all outer space.
About the HUMAN-AI PROCESS
The song was composed by an AI model that is the result of Pedro's PhD project. This model, a Transformer-based model, was designed to generate tablatures, a type of musical notation that is most suited to guitar players, for it shows not only what to play, but also how to play it on the instrument (the first few seconds of the song are an actual playback of said tablatures within the Guitar Pro software). For this, the model was trained on a corpus of around 26k tablatures (https://github.com/dada-bots/dadaGP). These tablatures are first converted into a text-like format (i.e. the token format), and are then fed to the model so that it learns the patterns between instruments, musical phrases, styles and all aspects relevant to composition. Due to the nature of the dataset (and the authors musical preference), its output focuses mostly on songs from the Rock and Metal sphere. The model generates not only guitar parts, but all the instruments heard in the song. In terms of iterations, a curation process was undertaken by Pedro and Jack, selecting some compositions from the model that were generated across the years of Pedro's PhD. With that selection done, the pieces were imported into a DAW and merged together to create a coherent Metal song. Guitars and bass guitar were recorded, and drums were sonified using VSTs. Finally, a mixing step was done by Jack.