Participants 2022
Stochastic Monkeys
AI Song Contest 2022 / participantsTEAM / Stochastic Monkeys
SONG / Talkoholism
TEAM MEMBERS / András Barják, Gabriel Vigliensoni
About the TEAM
Gabriel Vigliensoni is a Montréal-based electronic music artist, performer, and researcher whose work interrogates the different stages of contemporary music production’s workflow, always transforming the process of making a record into a playing field for experimentation and learning. Having studied sound recording and music technology in Santiago and Montréal, and carried out research on machine learning for creative practice in London, vigliensoni is equally grounded on the current state-of-the-art and experimental techniques for music-making as well as on the electronic music subcultures surrounding house and techno.
András Barják is a Stockholm-based audio software developer (at XLN Audio), ML engineer, and classically trained multi-instrumentalist musician who enjoys creating music in a wide variety of genres. Having spent 10+ years in Hungary as a professional sound engineer and music teacher he understands music production and music theory well. Around 2017, one day, András woke up with a powerful dream in which he was using an AI-powered plugin to generate realistic flute performances from MIDI. This prompted him to switch careers and focus on ML-based audio software development.
About the SONG
Talkoholism explores different flavors of collaborating with AI: mapping gestures to control timbre exploration, generating musical events from rhythmic models learned from custom datasets, modeling audio from a contributed dataset of South American audio beauties and, finally, being accompanied by a text-based music composition expert assistant.
About the HUMAN-AI PROCESS
Our vision was to generate creative seeds using AI tools for promoting and advancing the different music production stages of a song. This technique is akin to Oblique Strategies (1975), a set of cards by Brian Eno and Peter Schmidt with the goal of unblocking makers in their everyday creative practice. The core concept was created by prompt engineering OpenAI’s GPT-3 (v. da-vinci-002) into a music composer expert assistant (MCEA). This interaction produced a few boring interactions, a couple of hilarious discussions, but in the end we figured out the right way to steer the discussion for more experimental ideas. We collected ideas about creative concepts, song structure, orchestration, effect usage, and melodic excerpts. In parallel, we investigated the creative affordances of state-of-the-art realtime neural audio generation. We set up an instance of RAVE (Caillon and Esling. 2021) and trained it on a custom-built audio dataset. This dataset consists of 4.5 hours of South American indigenous chants, poetry, testimonies, audio documents about space and time, and radio signals. Later in the orchestration and production processes, we brought in various AI-generated beats, AI-driven instruments, and AI-driven FXs using tools such as OpenAI’s Codex, R-VAE, DDSP and 3rd-party commercial smart plugins such as Sonible’s. We took the core ideas proposed, the melodies and sounds, and built upon and transformed those with our human composition skills. We experienced and unfolded a unique creative process that neither of us would have created on our own—without each other—and without the AI.