Design and Evaluation of Noise Reduction Techniques for Binaural Hearing Aids (Ontwerp en evaluatie van ruisonderdrukkingstechnieken voor binaurale hoorapparaten)
Author:
Keywords:
SISTA
Abstract:
One of the main complaints of hearing aid users is their degraded speech understanding in noisy environments. Modern hearing aids therefore include noise reduction techniques. These techniques are typically designed for a monaural application, i.e. in a single device. However, the majority of hearing aid users currently have hearing aids at both ears in a so-called bilateral fitting, as it is widely accepted that this leads to a better speech understanding and user satisfaction. Unfortunately, the independent signal processing (in particular the noise reduction) in a bilateral fitting can destroy the so-called binaural cues, namely the interaural time and level differences (ITDs and ILDs) which are used to localize sound sources in the horizontal plane. A recent technological advance are so-called binaural hearing aids, where a wireless link allows for the exchange of data (or even microphone signals) between the two devices. An advantage is that the signal processing can then be coordinated so that the binaural cues are better preserved.The goal of this thesis work is to develop a binaural noise reduction technique for binaural hearing aids. The binaural technique should achieve further speech intelligibility improvements (compared to a bilateral technique), while the binaural cues of both the speech and the residual noise should also be better preserved. For this purpose, the Speech-Distortion-Weighted Multi-channel Wiener Filter (SDW-MWF), which is a well-known multi-microphone noise reduction technique, is adopted in this thesis. It will be demonstrated that the SDW-MWF is particularly suited for binaural noise reduction, both in terms of noise reduction performance and in terms of binaural cue preservation.The first part of the thesis gives a general introduction to the problem statement. A theoretical binaural noise reduction framework is then presented. Finally, the binaural SDW-MWF technique is introduced. The second part of the thesis focuses on the noise reduction performance of the considered binaural SDW-MWF technique. A perceptual evaluation of a realistic adaptive implementation of the binaural SDW-MWF is presented first. It is demonstrated that the binaural SDW-MWF indeed allows for significant speech intelligibility improvements (compared to bilateral techniques), both for normal hearing subjects and for hearing aid users, which motivates further development and optimizations of the technique. A theoretical analysis of new SDW-MWF formulations is then presented. It is demonstrated that implementations based on these expressions can lead to further improvements in noise reduction performance (compared to the previously evaluated implementation). It is also demonstrated that the conceptually interesting structure of these new SDW-MWF formulations can be exploited in a number of ways. First, it is demonstrated how the structure can be used to increase the robustness to certain estimation errors, which occur in practice. Second, it is demonstrated that the structure also allows for a numerically favorable implementation.The third part of the thesis focuses on the binaural cue preservation of the considered binaural SDW-MWF technique. First, it is proven that a full-bandwidth binaural SDW-MWF technique (where all microphone signals can be exchanged) preserves the binaural cues of the target speech source, while extensions can be formulated so that the residual noise cues can also be preserved. Second, we focus on the (speech) cue preservation of reduced-bandwidth binaural or bilateral SDW-MWF techniques (where not all microphone signals can be exchanged). It is then shown that the speech ITD cues are still preserved, whereas the speech ILD cues are distorted. Novel reduced-bandwidth techniques, which better preserve the speech ILD cues, are also proposed.The fourth part of the thesis presents a binaural Voice Activity Detection (VAD) algorithm, which is used to detect periods of speech activity. Many noise reduction techniques (including the SDW-MWF) rely on a VAD, hence a robust VAD algorithm is highly sought for. The approach here consists in utilizing the wireless link of the binaural hearing aids to combine local VAD decisions into a superior fused VAD decision. Two novel binaural fusion VADs, which indeed achieve an improved performance versus the local VAD algorithms, are proposed.