Generalization of Ego-noise Suppression Algorithms for Different Robot Exemplars

Proposal for a Bachelor Thesis


Generalization of Ego-noise Suppression Algorithms for Different Robot Exemplars


Ego-noise, i.e., the self-created movement noise of robots, reveals typical spectral and spatial patterns which are determined by a limited number of degrees of freedom. This motivates the use of learning-based approaches for ego-noise suppression, which is a key pre-processing tep in robot audition. So far, research was limited to egonoise created by a specific exemplar of a certain robot type. However, due to manufacturing tolerances, ego-noise created by different exemplars of the same robot type was shown to vary significantly. In this bachelor hesis, the performance degradation of common machine learning-based ego-noise suppression algorithms should be investigated if they are applied for ego-noise of a previously unseen robot exemplar.

For this, a library with ego-noise of different robots exemplars has been recorded. First, the student should apply different learning-based approaches for ego-noise suppression on the recordings of each robot exemplar individually. Subsequently, it is evaluated how the robot
xemplar-specific models generalize for ego-noise of the other robot exemplars. Based on this, it should be investigated how the generalization can be improved by training the models using more diverse data containing ego-noise of different robot exemplars.



Prof. Dr.-Ing. W. Kellermann


Alexander Schmidt, M.Sc., 05.021 (Cauerstr. 7),