Participants 2025

While writing the song, we looked for ways to incorporate deep learning models into the production in unconventional ways without taking away any of the creative process. Probably the most interesting/useful technique discovered was using a singing voice conversion model along with a sample of an explosion taken from youtube to change the timbre of the lead vocal to the timbre of the explosion. While this sounds pretty awful on its own, inserting it into the mix behind the lead vocal thickens up the lead vocal in the same way a recorded vocal double might, reducing the amount of vocals needed to be recorded. This same technique can also be used to create distinct harmonies and stereo vocal doubles. Another sound design element involved taking random gaussian noise and decoding it to audio using a pretrained audio autoencoder, resulting in an eerie and digital synth-like sound that was used in the intro and breakdown. Finally, we randomized the parameters of the Neural Amp Modeler, an open source amp modeling software which uses deep learning to learn profiles of amplifiers. This noisy amp profile was used to render the quad tracked guitar layers, allowing us to thicken up the chorus guitars in an interesting way. While these techniques were not necessarily flashy, they assisted the production in meaningful ways without getting in the way of creativity. Besides these elements, the song was written, recorded and mixed like a standard metal song.