Colloquium - Details
You will receive information about presentations in time if you subscribe to the newsletter of the Colloquium Communications Technology.
Master-Vortrag: Investigations on Supervised System Identification Algorithms
Montag, 13. Juni 2022
The system identification task aims at inferring the room impulse response of a specific acoustic enclosure. System identification is mandatory in applications such as acoustic echo cancellation and cross talk cancellation. Traditional gradient-based algorithms such as normalized least mean square algorithm uses FIR filters to estimate the RIRs, unfortunately, in a relatively high dimension. A novel dual-stage algorithm is proposed in this thesis. The algorithm performs a state update where allowed states are located on a manifold. In a first stage, an undercomplete autoencoder is trained over the RIR data set. In the second stage, we perform the system identification tasks. Here the problem is reformulated such that the latent state is updated instead of the full impulse response. The trained decoder is then exploited to transform the latent variables to a proper impulse response. Evaluation is made between the reconstructed RIR with reference to the true RIR.
In this thesis, at first, the simulation framework generates RIR data set. Then autoencoders with different layer setups are trained on the generated data set. The qualified autoencoders are employed in the inference stage to perform the system identification tasks. Two crucial parameters, i. e., the latent dimension size and updating step size of the manifold are investigated under different SNR conditions. It is demonstrated that under noisy conditions, the proposed method outperforms the traditional NLMS approach. Evaluation results also show that lower bottleneck size design benefits the system identification with adverse noise