Machine Learning-based Reconstruction of Relative Transfer Functions

Proposal for a Project Thesis


A Denoising Autoencoder for Speech Enhancement


The identification of a relative transfer function (RTF) between the signal components in two microphones evoked by a single acoustic source is an important
component of multichannel communication systems in noisy and reverberant
environments. However, especially in noisy environments or in case of weak
excitation of certain frequencies, the RTF identification exhibits poor accuracy.
To overcome such shortcomings, mathematical models can be used to reconstruct
the RTF from imperfect measurements. These include a priori learned geometric
structures of RTFs from the area of potential source positions, i.e., manifold
learning [1], and compressed sensing approaches for the reconstruction of the RTF
from incomplete measurements [2].
The aim of this thesis is the implementation and evaluation of supervised RTF
estimation algorithms which are able to determine RTFs from imperfect measurements starting with [1] and [2]. The implementation should be done in MATLAB.
As prerequisites, the student should have basic MATLAB programming experience
and an affinity to math.



Prof. Dr.-Ing. Walter Kellermann


M.Sc. Andreas Brendel, room 05.018,