Participants 2023
Berfoed
AI Song Contest 2023 / participantsTEAM / Berfoed
SONG / Revisit to the observatory
TEAM MEMBERS / Philip Bergström, Agnes Kofoed Christianson
About the TEAM
Philip and Agnes met while doing an improvisation project at the Academy of Music and Drama in Gothenburg. Philip is pursuing a degree in classical violin playing, but previously has a master’s degree in Machine Learning from KTH in Stockholm. He currently spends most of his time practicing, but also enjoys experimenting with music technology. Agnes is a singer and composer who creates interdisciplinary performances, experimenting with poetry and tonalities within different cultural contexts. Her works have been performed at several festivals such as the Oslo and Copenhagen Jazz Festival, as well as in the Stockholm Concert Hall.
About the SONG
Our song and its lyrics are based on the poem “Visit to the observatory” by the Swedish poet and nobel prize winner Harry Martinson (with permission from his family). It’s a song about the mysteries of the cosmos and its boundless potential. The piece harnesses the powerful creative possibilities of text-to-music models to directly transform content between the two artforms - written poetry and music.
Every 30 second segment in the piece corresponds to a stanza in the poem. It features an AI-generated voice as well as Agnes' improvisational song. This creates a contrast which aims to emphasize the difference between an artificial, imitating model and a living human.
About the HUMAN-AI PROCESS
The instrumental content of the song was generated by using the four stanzas of Martinson’s poem “Visit to the observatory” in English translation as inputs to a pre-trained text-to-music model (AudioCraft), generating four 30 second segments of music. The prompts were combined with the inputs: “a symphonic poem”, “G major” and “60 BPM”.
Using the verses of the poem, we produced four new stanzas using GPT-3.5 that were used as lyrics in our song. To create AI transcriptions of the new lyrics, we used a text-to-speech model (VITS) which was trained on the public VCTK dataset. This was contrasted by recording Agnes performing improvised singing and sounds in different layers, based on the lyrics.
Finally, we mixed all the tracks in Ableton. In order to mix the output from the AudioCraft model, we used Demucs, a model that is capable of performing music source separation by being trained on MUSDB, a dataset consisting of already isolated instrument stems.
Lyrics
In the light of our little tube at night, we dreamed of the eternal, the grand. In cosmic silence, beyond the grasp of time, we found ourselves lost, infinite, distant.
A golden mist swirls, a choir of stars in harmonious song, they weave threads of glowing light, and create images of the heavens' paths.
In this journey to other spheres, where each star is a book to read, where we become part of cosmic streams, the source of truth remains infinitely near.
So let us wander through the womb of space, explore worlds our minds create, and in the golden mist of our dream, find answers to questions we could never foresee.