When
Thematic Session 2: Software and Synthesis (Monday, 15:40)
Abstract
Deep learning models applied to raw audio have rapidly gained relevance in modeling and synthesis scenarios. These architectures have proved to be beneficial in cases where nonlinear phenomena are present. Nonlinearities have often a strong influence in the real world and they give unique tastes that are usually difficult to replicate. Nonlinear phenomena are usually described by complex equations, computationally expensive, difficult to solve, and sometimes difficult to formalize. Deep learning networks are universal approximators and this characteristic makes them appealing for black box approaches and so avoiding the effort of specific mathematical formulations. On the other hand, black box modeling can suffer from flexibility and interpretability, since no physical variables are considered and most of the time a static representation of the phenomena is captured in the modeling process.
However, recent advances have shown these techniques to be able to overcome these problems. Different approaches have been taken to include tunable parameters, using different ways to condition the networks to different dynamic scenarios and link the model variables to the physical properties of a real-world phenomenon using grey-box approaches to give interpretability to the model.
Finally, this presentation aims to give a brief review of the different architectures available so far for modeling nonlinear devices, but that can extend to other situation as embodied interactions.
Bio
Riccardo is a computer science engineer who graduated from the University of Padua. His interest is in audio modeling and synthesis, in particular, focusing on nonlinear phenomena that present acoustic and electronic instruments/devices. Now is pursuing a Ph.D. at the University of Oslo, where he is addressing nonlinear audio modeling using deep learning techniques. The research takes a specific acoustic instrument, the Pianoforte, and aims to model the nonlinearities included in its mechanism, often difficult to include in digital models. Therefore, the project aims to recreate a physical piano in the digital domain capable of bridging the gap between the analog and the digital world.