Colloquium - Details
You will receive information about presentations in time if you subscribe to the newsletter of the Colloquium Communications Technology.
Master-Presentation: Acquisition and Analysis of a Dataset for MIMO System Identification Using Manifolds
Mihaly Barany
Thursday, February 26, 2026
10:00 AM
IKS 4G | zoom
The task of tracking time-varying systems with multiple inputs and multiple outputs (MIMO) affects numerous applications in digital signal processing. Well-known applications are acoustic echo cancellation (AEC) or adaptive crosstalk cancellation. Usually, the task is solved with adaptive filters that do not use a priori knowledge about the tracked system. New approaches which use such knowledge during the system identification task, reach a faster and robuster system identification performance. In this case it is assumed, that all the possible impulse responses of a given acoustic scenario lie on a so-called low-dimensional manifold. Low-dimensional representations of the impulse responses can be determined by autoencoders. In order to learn such manifolds of a given acoustic scenario, a large amount of a priori knowledge, in this case room impulse responses, is required with high variability.
As part of this thesis, two data sets of MIMO impulse responses were recorded in cars in multiple different acoustic scenarios. In order to capture real-world impulse responses, the measurements were conducted with test persons sitting in the cars, who carried out small movements during the recording. The time-variant impulse responses were extracted from these measurements with a Kálmán Filter combined with Expectation Maximization. The captured impulse response datasets were investigated in the latent space with PCA and t-SNE. In the investigation and the comparison of the results it was found that the global parameters of the acoustic scenarios, such as window positions or seat occupations, have a significant effect on the impulse responses. In addition to the training datasets, two test datasets containing speech recordings were captured in a MIMO setup in both cars.
As the next step, linear manifolds were trained with PCA on the training dataset and the performance of an EKF was tested on the test dataset in an AEC scenario. Additionally, MIMO and MISO setups of the manifold Kálmán Filter were compared with each other. During these investigation it was found that by exploiting a priori knowledge, robuster performance can be achieved in time-variant acoustic system identification.
