Semi-blind Source Separation using Prelearned Demixing Filters

Proposal for a Bachelor Thesis / Research Internship


Semi-blind Source Separation using Prelearned Demixing Filters


In many daily-life scenarios, acoustic sources never can only be observed with interfering sources and noise. Hence, source separation and signal extraction are essential tasks for acoustic signal processing on a variety of devices such as mobile phones, smart home assistants,
earing aids, conference systems etc. Typically blind approaches such as Independent Component Analysis are used, which make only very generic assumptions about the problem and are hence applicable to a broad variety of problems.

However, due to the inherent generic assumptions, blind methods tend to fail in adverse scenarios. To improve the performance of blind source separation algorithms prior knowledge about the acoustic scenario can be introduced transforming the blind approach into a semi-blind approach. Such prior knowledge may consist of previously estimated signal cancellations filters, which are combined by an appropriate ICA algorithm to achieve interference suppression instead of blindly estimating the filter.

The aim of this thesis is the implementation and evaluation of acoustic source separation algorithms that take previously determined signal cancellation filters into account. Starting with [1], different selection and combination strategies of the previously determined cancellation filters should be investigated. As prerequisites, the student should have MATLAB programming experience and affinity to math.



Prof. Dr.-Ing. Walter Kellermann


M.Sc. Andreas Brendel, room 05.018,


Ab sofort