Master Theses - Details

Investigations on Supervised System Identification Algorithms

Betreuer:Tobias Kabzinski, Till Hardenbicker

Themengebiet: machine learning, signal processing, adaptive filters

Kategorie: Masterarbeit (MA)

Status: Offene Arbeit

Tools: Python, Matlab

Task description

The system identification problem is encountered in many digital signal processing areas. Prominent applications are the acoustic echo control (AEC) problem or the continuous measurement of head-related transfer functions (HRTFs). While today's approaches, such as the normalized least-mean-square (NLMS) algorithm, do not exploit a priori knowledge about the system, exploiting this prior knowledge can potentially bring about advantages in convergence speed or in tracking of time-variant systems. This way, potentially more precise or faster system identification could be achieved.

In this thesis, fundamental investigations on the machine learning-based representation of room impulse responses, binaural room impulse responses or HRTFs shall be conducted. Building on these results, a machine learning-based extension of classical adaptive filtering algorithm for the impulse response identification shall be developed and compared against (simple) state-of-the-art algorithms.

One or more of the following aspects shall be investigated:

  • investigations on manifold learning for a low-dimensional representation of impulse responses, e.g., using an auto encoder architecture
  • investigations on using a neural networks to represent classical signal processing methods, such as deconvolution
  • investigations on manifold learning for the delay estimation problem (with synthesized data)
  • development of adaptation algorithms in the learned low-dimensional space, for instance, based on manifold learning
  • extension of methods to cope with single-input multiple-ouput (SIMO) and/or multiple-input multiple-output (MIMO) systems

If you are interested and would like to know more, do not hesitate to get in touch with us.