Start Simple! Learning Sound-Producing Systems from Surrogate Problems

When: Wednesday, July 29, 2026, 3:00 PM - 4:00 PM
Where: Engineering Building, G2, Mile End

Abstract: Modern generative models have transformed audio synthesis, yet they are primarily designed to reproduce signals rather than uncover the physical systems that generate them. This talk presents an alternative perspective on learning from sound, drawing on ideas from classical system identification and recent advances in representation learning. Rather than fitting increasingly complex models directly, it advocates for progressively learning through a sequence of surrogate problems that incorporate physical insight and structural priors. This perspective brings together generative modelling, differentiable signal processing, and a homotopy-based formulation of system identification. The goal is not simply to generate more realistic sounds, but to develop interpretable models that reveal the mechanisms, dynamics, and structure of the systems that produce them.

Bio: Prof. José Luis Blanco currently serves as Head of the Signal Processing Applications Group (GAPS) and Deputy Head of the Department of Signals, Systems and Radiocommunications (SSR) at Universidad Politécnica de Madrid. He holds a Telecommunication Engineering degree from UPM, 2007, masters in Signal Theory, 2009, and the PhD in Digital Signal Processing and Machine Learning, 2013, with international research stays at Oxford University and Université Libre de Bruxelles. His research lies at the intersection of neural analysis-by-synthesis, differentiable optimization, and statistical digital signal processing. His work focuses on the development of optimization-driven computational frameworks that combine informed models with learning techniques to characterize, synthesize, identify, and compensate the mechanisms underlying complex signals and their production mechanisms.