Robustness Analysis of Binaural Hearing Aid Beamformer Algorithms by Means of Objective Perceptual Quality Measures

Abstract

In this contribution different microphone array-based noise reduction schemes for hearing aids are suggested and compared in terms of their performance, signal quality and robustness against model errors. The algorithms all have binaural output and are evaluated using objective perceptual quality measures [1,2,3]. It has been shown earlier that these measures are able to predict subjective data that is relevant for the assessment of noise reduction algorithms. The quality measures showed clearly that fixed beamformers designed with head models were relatively robust against steering errors whereas for the adaptive beamformers tested in this study the robustness was limited and the benefit due to higher noise reduction depended on the noise scenario and the reliability of a direction of arrival estimation. Furthermore, binaural cue distortions introduced by the different binaural output strategies could be identified by the binaural speech intelligibility measure [3] even in case monaural quality values were similar. Thus, this perceptual quality measure seems to be suitable to discover the benefit that the listener might have from the effect of spatial unmasking.

Objective perceptual quality assessment for self-steering binaural hearing aid microphone arrays

“…PSM / ΔPSM: The quality measure PSM from PEMO-Q [7] estimates the perceptual similarity between the processed signal and the clean speech source signal. It has shown high correlations between objective and subjective data and has been used for quality assessment of noise reduction schemes in [1,8,9]. PSM increases with increasing (input) SNR.…”

“…The fixed noise-field characteristic is coded in the coherence matrix ΓNN[m] which additionally influences beamformer properties directivity and susceptibility to white noise, and therefore has to be constrained [4,1]. Both, d[m, Θ] and ΓNN[m] depend on to the assumed wave propagation model which may differ from the true (and generally unknown) wave propagation from the source to the microphones.…”

“…For headworn arrays it is beneficial to include knowledge about head shadow and diffraction effects [1,11], especially for lateral target signal sources. Thus, head models by Duda et al [5,6] are applied which are effective parametric models that are based on the characteristics of a sphere.…”

“…However, as the speech intelligibility is a nonlinear function of the SNR and other signal features such as the preservation of binaural cues, we use the difference between output and input SRT, namely the ΔSRT, as an indirect measure for the increase of intelligibility. The binaural SRT measure as described in [10,1] assumes a spatially stationary source configuration. To be applicable to moving sources it had to be extended to a block-wise measure with subsequent averaging across blocks.…”

“…This may become unsatisfying for the hearing aid user in particular for algorithms with a high spatial selectivity and if the signal of interest is moving. In this contribution a combination of a binaural beamformer [1,2] and an automatic steering (electronic control of the look direction) based on the Generalized Cross Correlation (GCC) approach by Knapp and Carter [3] is applied. The importance of a proper model of wave propagation is investigated for a head-worn DOA-beamformer system.…”