|Supervisor:||Dr.-Ing. Klaus Reindl (Room 5.13)|
|Faculty:||Prof. Dr.-Ing. Walter Kellermann|
Noise reduction algorithms are crucial for hearing aid applications. The applied algorithms should significantly reduce noise and interference components and at the same time they should not introduce audible artifacts. These conflictive goals make it difficult to realize noise reduction algorithms for hearing aids.
There are numerous noise reduction algorithms in literature where spectral subtraction is known best and one of the simplest strategies. On the other hand, it is also known as the concept which leads to most signal distortion. However, noise reduction techniques are often based on the short-time Fourier transform (STFT) which may not be appropriate for speech signal processing and hearing aid applications. It is well-known that the frequency resolution of human hearing is nonuniform and described by critical bands or a bark scale. The fact that the signal processing strategy does not match human perception may explain the audible distortion of noise reduction techniques.
In this thesis, the noise reduction concept for binaural hearing aids developed at the LMS shall be implemented with a gammatone filter bank which models the frequency resolution of human hearing. A theoretical as well as an experimental investigation by Matlab simulations is necessary for different multispeaker, noisy, and reverberant conditions. The gammatone filter bank implementation shall be compared to the current polyphase filter bank implementation by analyzing speech distortion and the SINR gain at the outputs of both schemes. Additionally, for a comprehensive evaluation of both implementations, the output signals should be analyzed by using the PEASS toolkit which provides objective measures for perceptual evaluation of separated audio signals [Emiya et al. 2011]. Well-documented and well-structured software is important. The thesis can be written in German or English