Abstract

Modern hearing aids feature multiple microphones that are suitable to conduct directional filtering, e.g., beamforming. Speech enhancement is an important step in signal processing conducted in such multi-channel assistive hearing systems and it requires robust localization of the target speaker as well as proper spatial filters. Spatial information about target as well as interfering sound sources is therefore a basic requirement. Filter parameters can be determined based on assumptions like free-field sound propagation and omni-directional microphone characteristics. This allows a parametric filter design based on the direction of arrival of a sound source, but does not hold to the whole extent in reverberant environments and for real sensor setups. An alternative is the estimation of sound source-related parameters adaptively from the input signal. I will talk about my recent work on the use of probabilistic spatial information of sound source locations for the estimation of these parameters. Spatial information is obtained from a classifier-based localization method that is trained on the acoustic characteristics of the human head and, secondly, on a multi-channel setup from the CHiME-3 speech recognition challenge. I present results in terms of speech enhancement in hearing aids as well as speech recognition performance on CHiME-3 data.

https://www.uni-oldenburg.de/mediphysik-akustik/mediphysik/mitarbeiter/hendrik-kayser/