���ʵ���

Title of the thesis: Domain-Aware Deep Learning for Room Acoustics: Parameter Estimation, Localization, and Source Separation

Thesis defender: Ricardo Falcon Perez
Opponent: Prof. Juan Bello, New York University, US  
Custos: Prof. Ville Pulkki, Aalto University School of Electrical Engineering 

Sound is shaped by the spaces it travels through. The same signal can be clear, muffled, or immersive depending on room acoustics. And this can affect both people and technologies that listen to, analyze, or reproduce sound.

This doctoral thesis explores how machine learning can work together with acoustic and signal-processing knowledge to understand and use these effects, for both real-world environment as well as virtual spaces. The research therefore develops computational methods that can predict or handle acoustic behaviour without requiring extensive physical measurements.

The thesis organizes its contributions around three ways acoustics appears in machine learning. First, acoustics as the target, where the goal is to estimate acoustic properties of spaces from available data. Second, acoustics as interference, where room effects make listening tasks harder and models need to become more robust. Third, acoustics as both a challenge and an opportunity, where difficult conditions like reverberation complicate tasks, but acoustic structure, such as spatial cues, can also provide useful guidance for learning.

The results include improved methods for estimating room-acoustic characteristics, a technique that strengthens spatial sound detection by making models less sensitive to acoustic variation, and a framework that can separate overlapping machine sounds even when “clean” training examples are not available.

The main finding is that audio AI becomes more reliable when it is guided by the structure of acoustics, rather than relying on pattern recognition alone. The research brings new ways to combine learning with knowledge about how sound behaves in space—supporting more robust and meaningful analysis of real acoustic environments.

These insights can be applied in spatial audio and immersive media, smarter microphones and machine listening, acoustic scene analysis, and monitoring of machines in real environments. The thesis concludes that future progress will require not only larger datasets and models, but also approaches that reflect how sound propagation and human perception shape what we hear.

Key words: Room Acoustics, Sound Source Separation, Sound Event Localization and Detection, Acoustic Parameter Estimation

Thesis available for public display 7 days prior to the defence at .

Contact: