EverSong - singing voice transfer with Google Magenta’s DDSP + diffusion-based vocoder. (emf2024)

Emily Peasgood and Mark Hanslip will present a talk on creating EverSong, followed by an invitation to try their vocal transfer models in real-time. EverSong is a sound installation where you can hear the lifespan of a person in a song, from 5-years-old at the start to 95-years-old at the end. It explores the fragility of life, its circularity, invites people to celebrate the moment we are in right now and questions our legacy as humans moving into an increasingly digital age. Since January 2024, Peasgood & Hanslip have worked with DDSP-SVC, a newly-developed process which builds on Google Magenta’s DDSP by combining it with a diffusion-based vocoder. They recorded vocal datasets with people aged 5-95 and trained models which can be mapped onto Emily’s voice enabling her to sound younger and older than she is. This technology is starting to be adopted by contemporary pop musicians, high-profile examples being web services such as Grimes’ Elf and Holly Herndon’s Holly+. EverSong represents a new application of this technology by making it possible to hear a voice age in real time. It will be experienced in tunnels and pathways, installed sustainably in nature, touring before finding a permanent home in an ancient site. about this event: https://www.emfcamp.org/schedule/2024/120-eversong-singing-voice-transfer