Termin – Detailansicht
Master-Vortrag: Identification of Time-Variant MIMO Systems using Acoustic Manifolds
Johannes Hahn
Montag, 14. Oktober 2024
14:00 Uhr
IKS 4G | hybrid
Many applications in digital signal processing such as Acoustic Echo Cancellation (AEC) and adaptive Crosstalk Cancellation (CTC) rely on methods of Acoustic System Identification (ASI), which aims at determining the impulse responses between loudspeakers and microphones in a given acoustic enclosure. This task is particularly challenging if multiple inputs and outputs have to be considered, leading to a large number of parameters to be identified. The complexity increases further if the system and thus the parameters are time-varying. Moreover, such Multiple Inputs Multiple Outputs (MIMO) systems suffer from the non-uniqueness problem, which impairs the performance of established ASI methods. To overcome these challenges, recent developments focus on the incorporation of prior knowledge into the system identification. In this context, several approaches take advantage of the manifold hypothesis. It states that the Room Impulse Responses (RIRs) in a given MIMO scenario can be modeled by a low-dimensional representation lying on or close to a usually nonlinear manifold, which can be learned a-priori or estimated online.
This thesis investigates a novel approach to the identification of time-variant MIMO systems which exploits the manifold hypothesis to greatly reduce the number of tracked parameters. A Variational Auto-Encoder (VAE) is trained to learn the manifold in terms of a low-dimensional coordinate representation. During ASI, only this representation is tracked, using an Extended Kalman Filter (EKF). As a result, less time-variant parameters have to be identified.
Additionally, measurements are conducted using an artificial head and two loudspeakers to create a data set of real-world RIRs, and test sequences are recorded that facilitate the evaluation of ASI algorithms in a real-world scenario. In an extensive parameter study, the architecture and hyperparameters of the VAE are investigated and tuned. A validation of the proposed method in simulated scenarios as well as on the test recordings showcases its superiority in the identification of time-varying systems and its robustness towards the non-uniqueness problem in comparison to established ASI methods and related approaches built on the manifold hypothesis.
