Mobile robots create noise while moving. This noise, which is referred to as ego-noise, strongly disturbs the microphone recordings of the robot, which are used to interact with the acoustic environment. To reduce or suppress the ego-noise in these recordings, the use of Deep Attractor Networks shall be investigated.
The Deep Attractor Network is a single-channel deep neural network approach to source separation. It builds on Deep Clustering, which is currently a hot topic in source separation. The general idea of both setups is to embed the time-frequency representation of a mixture in a higher dimensional space and perform source separation in this embedding domain. A main contribution of Deep Attractor Networks is that they can be trained in an end-to-end manner and hence promise to solve some of the problems that remain with Deep Clustering.
In this thesis, the suitability of Deep Attractor Networks for ego-noise reduction shall be investigated by implementing a Deep Attractor Network and evaluating its performance on already existing data. The results shall be compared to the performance of Deep Clustering and the advantages of one over the other should be evaluated.