Informed Sound Source Localization for Hearing Aid Applications

Mojtaba Farmani, Mojtaba Dissertation, Farmani, James Kates, Patrick Naylor, Mojtaba Farmani
2017 Aalborg Universitetsforlag. PhD Series   unpublished
Hearing impaired listeners often face difficulties in understanding speech, especially in noisy situations. A highly effective solution to solve this problem is to employ a hearing aid system (HAS), which can connect to a wireless microphone worn by the talker of interest. The wireless microphone allows the HAS to access an essentially noise-free version of the target signals that can be presented to the user. However, despite the increase in intelligibility, some users do not feel comfortable
more » ... ith this solution, because it does not provide the correct spatial cues of the target sound, so that the user cannot localize the target talker. This can reduce a user's sense of immersion and can cause the user to feel detached from the surroundings. Further, in situations where several talkers are simultaneously present and each of them are wearing a wireless microphone, lack of spatial cues can degrade the speech intelligibility of target signals, especially when some of the talkers are talking concurrently. One solution to address these problems is to impose the correct spatial cues on the wirelessly received signals, before rendering them to the HAS user. To do so, one could solve the informed sound source localization (SSL) problem, i.e estimate the location of the target talker(s) based on the knowledge of the noise-free version of the target signal(s). Despite the fact that the informed SSL problem is mainly relevant in acoustically noisy situations, and that HAS microphones are typically located behind/in the users' ears, existing informed SSL algorithms often ignore ambient noise characteristics and effects of the user's head on the received signals. In this thesis, we propose a maximum likelihood (ML) framework for solving the informed SSL problem, which allows to take both ambient noise characteristics and effects of a user's head into account. Ambient noise characteristics can be relatively easily estimated based on the wirelessly available noise-free target signal and the noisy target signals captured by the HAS microphones. To model effects of the head, we employ four different head models, which include generic models, which do not depend on a specific user, and individualizable models, which allow to take user-specific details into account. For each of the head models, we propose an informed SSL algorithm using the ML framework. Eventhough the computational complexity of the proposed methods differ, each of v Abstract the proposed algorithms has been formulated with computational efficiency in mind. Some of the proposed methods are flexible in the sense that they do not depend on a particular microphone array configuration. For these methods, we study how the microphone array geometry affects their performance. This is important because some microphone configurations (e.g. binaural) may require higher implementation costs than others (e.g. monaural). Finally, we assess the performance of the proposed methods in different noisy and reverberant conditions to demonstrate and to compare their effectiveness. vi