Adaptive filtering techniques for noise reduction and acoustic feedback cancellation in hearing aids

Understanding speech in noise and the occurrence of acoustic feedback belong to the major problems of current hearing aid users. Hence, an urgent demand exists for efficient and well-working digital signal processing algorithms that offer a solution to these issues. In this thesis we develop adaptive filtering techniques for noise reduction and acoustic feedback cancellation. Thanks to the availability of low power digital signal processors, these algorithms can be integrated in a hearing aid. Because of the ongoing miniaturization in the hearing aid industry and the growing tendency towards multi-microphone hearing aids, robustness against imperfections such as microphone mismatch, has become a major issue in the design of a noise reduction algorithm. In this thesis we propose multimicrophone noise reduction techniques that are based on multi-channel Wiener filtering (MWF). Theoretical and experimental analysis demonstrate that these MWF-based techniques are less sensitive to imperfections than the widely studied Generalized Sidelobe Canceller (GSC). Thanks to their high robustness, they can fully exploit the benefit of an additional third microphone in a behind the-ear hearing aid and hence reduce more noise. In addition, we show that a subband implementation further improves speech intelligibility. Finally, we establish a generalized noise reduction scheme, called the spatially pre-processed speech distortion weighted MWF, that encompasses the GSC and the MWF as special cases and allows for attractive in-between solutions. Low-cost subband and stochastic gradient implementations are proposed that make the implementation of the developed techniques feasible in future commercial hearing aids. With the growing use of open fittings and the decreasing distance between the loudspeaker and the microphone of the hearing aid, the demand for effective feedback suppression techniques increases. A promising solution for acoustic feedback is adaptive feedback cancellation. However, the adaptive modelling problem encountered here appears to be highly non-trivial, because of the presence of a closed signal loop that introduces specific signal correlation. As a result, standard adaptive feedback cancellation techniques fail to provide a reliable feedback path estimate, when the desired signal is spectrally colored. In this thesis we develop adaptive feedback cancellation techniques that are based on closed-loop identification of the feedback path as well as an autoregressive modelling of the desired signal. Simulations show that the developed techniques provide a better feedback path estimate for spectrally colored and highly time-varying sound signals (such as speech) than the standard adaptive feedback cancellation techniques.