Colloquium - Details
You will receive information about presentations in time if you subscribe to the newsletter of the Colloquium Communications Technology.
Master-Presentation: Correction of Measurement Errors in Higher-Order Ambisonics Signals Using Machine Learning
Helena Janning
Thursday, October 30, 2025
03:00 PM
IKS 4G | zoom
The spatial recording and immersive reproduction of sound fields have gained significant popularity over the past decades, with a wide range of applications such as Virtual Reality (VR) and teleconferencing. In practice, the recording is usually realized by capturing a sound field with a spherical microphone array at discrete sampling points in space. Based on these recordings, the directional information of the sound field can be represented in the higher-order Ambisonics (HOA) sound format, which abstracts the sound field from the array geometry.
Although spherical microphone arrays are very useful in practical applications, they introduce measurement errors, which are caused by several factors such as spatial aliasing, non-ideal positioning of the microphones and self-noise of the microphones. In the literature, different approaches exist to mitigate the effect of aliasing errors. Conventional approaches aim to suppress spatial aliasing by formulating and solving an optimization problem. In addition, existing Deep Learning (DL) approaches take recorded signals from spherical microphone arrays as an input with the objective of estimating the ideal HOA signals.
This thesis proposes a novel DL approach that directly operates in the HOA domain. In the thesis, two different approaches are suggested: The first approach relies on theoretical simulations, which enable the training and assessment of the model on the isolated error types. The second approach aims to address more realistic applications by considering microphone arrays simulated within different rooms. Simulation experiments demonstrated that the model enhances the median spatial similarity to the target signal by 268%, outperforming conventional approaches that achieved an improvement of 106%. While the first approach exhibited limited generalization capabilities on realistic HOA signals obtained from spherical microphone array recordings, the second approach effectively achieved a reduction in aliasing and positioning errors in realistic recordings. Here, the median spatial similarity to the target signal is enhanced by 21.19%. These finding provide insights into the relationship between ideal and measurement-based HOA signals and demonstrate the capability of DL approaches to reduce measurement errors.
