C4DM seminar: Diffusion Models for Audio Effects: From Blind Estimation to Automatic Mixing

Topic: Diffusion Models for Audio Effects: From Blind Estimation to Automatic Mixing

Location: GC203, Graduate Centre, Mile End Campus, QMUL

This talk presents a series of recent works exploring the use of diffusion models in the analysis and generation of audio effects. In the first part, I will focus on blind and unsupervised effect estimation, where diffusion models are employed as powerful data-driven priors for recovering clean or unprocessed signals from their altered counterparts, while simultaneously estimating the parameters of a model of the underlying audio effect. In the second part, I will discuss a recent work on automatic music mixing, introducing MEGAMI (Multitrack Embedding Generative Auto MIxing)—a generative framework that models the distribution of professional mixes directly in a multitrack effect embedding space.

Eloi Moliner received his Ph.D. degree from the Acoustics Lab of Aalto University, Espoo, Finland, in 2025. He previuosly obtained his M.Sc. degree in Telecommunications Engineering in 2021 and his B.Sc. degree in Telecommunications Technologies and Services Engineering in 2018, both from the Polytechnic University of Catalonia, Spain. He received the Best Student Paper Awards at IEEE ICASSP 2023, IWAENC 2024, and AES AIMLA 2025. His research interests include generative models for audio, audio restoration and enhancement, and audio signal processing.

Webpage | Teams link 

Key contact: Haokun Tian