PhD defence by Xenofon Karakonstantis

Abstract

Sound is a complex phenomenon that fills our environments, whether we’re stepping into a virtual reality experience, sitting in a concert hall, or just in our living rooms. Understanding how sound behaves in these spaces is crucial for enhancing audio quality and creating immersive experiences. Yet, capturing and analyzing these sound fields poses significant challenges, especially when traditional methods meet their limits. The PhD thesis titled "Data-driven methods for large-scale sound field acquisition and analysis" addresses these issues head-on by harnessing the power of modern computational techniques and artificial intelligence.

The thesis explores how deep learning can overcome the limitations of conventional sound field analysis by developing advanced models that integrate the physics of sound with data-driven insights. For instance, it introduces generative models that learn from the statistical properties of sound fields, making them robust tools for reconstructing sound environments with high accuracy. These models are not just theoretical—they have practical applications in designing better acoustics for architecture and improving noise control, enhancing how we interact with sound in complex environments.

Furthermore, the research delves into physics-informed neural networks. These innovative networks embed the fundamental equations of sound directly into their architecture, allowing them to respect and utilize the underlying physics that governs sound propagation. This approach has proven effective in tasks like reconstructing room impulse responses and spatially detailed speech signals, paving the way for more accurate and efficient sound analysis tools.

Another significant advancement discussed in the thesis is the application of convolutional neural networks with rotational equivariance, tailored for spherical microphone arrays. This method significantly improves the localization of sound sources, crucial for applications in virtual reality and smart audio environments.

The potential uses of these data-driven approaches are vast. They promise to enhance virtual reality experiences with more realistic soundscapes, improve the acoustics of public and private spaces, and advance our general understanding of how sound interacts with its surroundings. This PhD work not only pushes the boundaries of acoustic science but also offers new tools and methodologies that could transform how we experience sound in our daily lives and in professional settings.

Supervisors

• Main supervisor: Associate Professor Efren Fernandez Grande, DTU Electro, Denmark
• Co-supervisor: Adjunct Professor Peter Gerstoft, Scripps Institution of Oceanography, University of California.

Assessment committee

• Associate Professor Finn Thomas Agerkvist, DTU Electro, Denmark (chair)
• Professor Toon van Waterschoot, KU Leuven, Belgium
• Associate Professor Sebastian Schlecht, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Germany

Master of the Ceremony

• Associate Professor, Cheol-Ho Jeong, DTU Electro, Denmark