4 novembre 2026 à 15h
Neural instrument synthesis offers new possibilities for music creation and production, yet developing high-quality, expressive, and controllable virtual instruments remains challenging. In this talk, I explore two directions in neural instrument synthesis: expressive performance rendering and instrument timbre cloning.
First, I introduce VIOLET, a latent-diffusion framework for violin synthesis that generates high-quality performances conditioned on MIDI notes, playing techniques, and continuous dynamics, enabling fine-grained control over musical articulation and expression. Second, I present Point-and-Play, a framework for instrument timbre cloning directly from polyphonic music mixtures. Using text or audio cues, the system extracts the timbre of a target instrument without explicit source separation and synthesizes new notes across different pitches and velocities.
Together, these works illustrate how neural synthesis can enable expressive performance control and flexible timbre reuse in music production.
Lien visio: https://visio.numerique.gouv.fr/alg-omus-ada
Room Agora 2
français
English