Subjective evaluation of signal-dependent partial noise estimation algorithms for binaural hearing aids

Speech understanding is still a major challenge for many hearing aid users in complex acoustic scenes with background noise and interfering speakers. Besides suppressing undesired sound sources, an important objective of a binaural noise reduction algorithm is to preserve the spatial impression of the acoustic scene for the hearing aid user. Although the binaural minimum variance distortionless response (BMVDR) beamformer results in a good noise reduction performance and preserves the binaural cues of the target speaker, it distorts the binaural cues of the background noise, such that the target speaker and the background noise are perceived as coming from the same direction. Aiming at also preserving the binaural cues of the background noise, the binaural MVDR beamformer with partial noise estimation (BMVDR-N) was proposed, where a parameter allows to trade off noise reduction and binaural cue preservation of the background noise. In [1] a method was proposed to determine this trade-off parameter based on psychoacoustically-motivated boundaries for the desired interaural coherence of the background noise component.

In this contribution, we present a signal-dependent method to determine the trade-off parameter for the BMVDR-N beamformer based on the coherence between the noisy input signals and the output signals of the BMVDR beamformer [2]. We compare the performance of the BMVDR-N beamformer (both trade-off parameter methods) with the BMVDR beamformer for a realistic acoustic scenario with a target speaker, diffuse babble noise and an interfering speaker. Speech intelligibility was evaluated using the Oldenburg sentence test and spatial quality was evaluated using a MUSHRA test with N=11 normal-hearing subjects. In the presence of only diffuse babble noise, the signal-dependent trade-off parameter for the BMVDR-N beamformer significantly improves spatial quality compared to the BMVDR beamformer, while achieving the same speech intelligibility. Compared to the method in [1], the presented method achieves a significantly better performance, both in terms of speech intelligibility and spatial quality. When in addition an interfering speaker is present, the BMVDR-N beamformer (both trade-off parameter methods) significantly improves spatial quality without decreasing speech intelligibility compared to the BMVDR beamformer.

[1] D. Marquardt, S. Doclo (2018). “Interaural Coherence Preservation in Binaural Hearing Aids using Partial Noise Estimation and Spectral Postfiltering,” IEEE/ACM Trans. Audio, Speech and Language Processing, 26(7):1257-1270.
[2] J. Klug, D. Marquardt, N. Gößling, S. Doclo (2018). “Evaluation of Signal-Dependent Partial Noise Estimation Algorithms for Binaural Hearing Aids,” in Proc. ITG Conference on Speech Communication, Oldenburg, Germany, Oct. 2018, pp. 236-240.