Reconstructing sound with a Bayesian approach
Patrick Wijnings defended his PhD thesis at the Department of Electrical Engineering on April 9.
Acoustic measurements using a multitude of microphones have many valuable applications, such as acoustic surveillance to reduce public security costs and acoustic product design to limit noise pollution. However, reconstructing the behavior of acoustic sources from microphone array measurements is challenging. As sound travels from its sources to the microphones, information is typically lost, leading to a difficult inverse problem. In his PhD research, Patrick Wijnings approaches this problem from a Bayesian perspective. He addresses the inverse problem in a principled way by modelling both the measurement process and additional information as random variables.
’ research branches into two main directions. The first is an extension that enables the automatic estimation of the signal-to-noise ratio in broadband measurements taken in the acoustic near field. This allows an automatic balancing between reconstruction resolution and sensitivity to noise. The method was first applied to live hummingbirds to investigate how oscillating aerodynamic forces shape the timbre of their hum. Then, it was validated by comparing the reconstructed estimates with predictions made by a simple, first-principles model of the 3D oscillating forces.
Want to know more about the research with live hummingsbirds? Watch .
Far-field acoustic source localization
The second direction is an extension for handling the non-linear transfer functions that arise in far-field acoustic source localization. This method robustly accommodates multiple sources located at varying distances from the microphone array and can incorporate geometric information about the sources' environment. This was validated by reconstructing some of the scenarios from the Benchmark for Room Acoustical Simulation in an anechoic chamber.
Collecting data
Across all experiments, Wijnings used arrays of micro-electro-mechanical (MEMS) microphones. To build confidence in their acoustic performance, he collected a large dataset containing the sensitivity and phase characteristics of more than eight thousand individual microphones. From this dataset, probability distributions were extracted, enabling the use of this information as a priori input to further enhance acoustic localization and quantification algorithms. In addition, Wijnings explored related aspects of efficient computation in his research.
Novel acoustic algorithms
This research has resulted in novel acoustic algorithms suitable for two real-world use cases. First, Wijnings developed near-field acoustical holography methods that operate on transient broadband signals. This is useful for acoustic product design. He also developed beamforming techniques capable of handling acoustic sources at various distances from microphone array(s) and incorporating a priori knowledge of the geometry of the sources’ environment. This is useful for acoustic intelligence applications.
Title of PhD thesis: . Supervisors: Prof. Henk Corporaal, Prof. Bert de Vries and Dr. Sander Stuijk.