Dominique FOURER intervient dans le Workshop ARM 2026, Évry, le lundi 20 juillet 2026.
Title
Deep Bayesian Inference for Audio Processor Parameter Estimation
Abstract
Many audio inverse problems are inherently ill-posed: several combinations of processor parameters may produce perceptually indistinguishable sounds. Consequently, estimating a single parameter configuration is often insufficient to characterize the solution space. In this talk, we present a Bayesian framework that estimates posterior distributions over audio processor parameters rather than deterministic values. The proposed approach combines a differentiable digital signal processing (DDSP) model of the audio processor with a Variational Autoencoder and Normalizing Flows to learn expressive, potentially multi-modal posterior distributions conditioned on the target sound. By leveraging differentiable synthesis, the framework can be trained end-to-end while naturally capturing the uncertainty and ambiguity of the inverse problem. The approach is illustrated on the estimation of audio effect parameters and highlights how deep generative models and DDSP provide practical tools for uncertainty-aware inference in audio signal processing.
- Date: Lundi 20 juillet 2026, 11h40-12h
- Lieu : Université Évry Paris-Saclay – Bâtiment Maupertuis – salle 02E14
- Page Web du Workshop ARM 2026
- Site Web de l’axe transversal IA (Dominique Fourer)
