Fast Multichannel Nonnegative Matrix Factorization (MNMF) for Ego-noise Suppression

Proposal for a Master Thesis


Fast Multichannel Nonnegative Matrix Factorization (MNMF) for Ego-noise Suppression


Ego-noise, i.e., the mechanical noise of robots, is a severe problem in robot audition since it corrupts recordings and therefore degrades performance of, e.g., a speech recognizer. To relieve this problem, a suitable noise reduction mechanism is required. Ego-noise is strongly structured both spectrally and spatially such that Nonnegative Matrix Factorization (NMF) is well suited to represent and approximate ego-noise.
NMF factorizes a non-negative data matrix into a so-called non-negative dictionary matrix of reoccurring spectral patterns in the data and non-negative activation coefficients that combine the dictionary elements to approximate the entries of the data matrix. In multichannel NMF (MNMF), this concept is extended by spatial characteristics associated to each entry in the dictionary allowing both a spectral and spatial modelling of the data.

A major challenge of MNMF is its high computational complexity. To relieve this problem, fast approximative algorithms have been proposed which, however, are not guaranteed to converge. Recently, an alternative approach has been proposed which is not only efficient but also guarantees convergence. In this thesis, this approach should be investigated and implemented in Matlab. The student should evaluate the approach thoroughly and compare its performance and computational complexity to previous MNMF implementations.


Prof. Dr.-Ing. Walter Kellermann


Alexander Schmidt, M.Sc. – – +49 9131 85-27108