All Answers Forbidden: From RAVE Models to a Virtual Singer by Bengisu Önder

  • Published on 18 fév. 2026 by Bengisu
neural synthesis RAVE virtual singer

The exploration of AI systems in artistic practice raises fundamental questions of intention and control. This presentation examines how the strengths of neural networks such as RAVE can be shaped within a compositional framework to construct an expressive virtual voice.

➡️ This presentation is part of IRCAM Forum Workshops Paris / Enghien-les-Bains March 2026

The presentation shows how the strengths and possibilities of neural networks like RAVE can be understood and shaped in order to reveal their poetic potential when guided by artistic intention toward an unusual and expressive result. All Answers Forbidden is conceived as a sonic space of interaction between two sides of a duality: a soprano and a column of five loudspeakers named Totem, representing her alter ego. The voice of Totem was built from RAVE models I trained on recordings of the soprano, along with a few public ones. By exporting versions at different training stages, I obtained sounds evolving from fragmented and unstable to more coherent and lyrical. This transformation becomes the expressive core of Totem. I use these instabilities to form a trajectory of becoming, tracing the voice’s growth from noise to lyric intensity and toward a kind of machine emotionality.