Authors: Moritz Bender¹, Giso Grimm¹, Volker Hohmann¹
¹CvO Universität Oldenburg
Background: Hearing aids aim to improve the daily experience of hearing impaired (HI) listeners with a focus on speech reception. They compensate for effects of hearing loss through amplification, dynamic range compression (DRC) and noise reduction. However, their performance in challenging environments with multiple talkers and background noise still has large potential for improvement. To this end, scene-aware DRC and neural DRC have been proposed. They aim at providing different types of DRC for segregated foreground speech and background noise components in a scene to enhance the perception of foreground speech. Nevertheless, they are limited by the performance of source separation between foreground and background signals. Also it remains unclear, which combination of foreground and background processing is most beneficial for HI listeners in real-life communication scenarios, in which speech is a common noise component.
Method: This study uses a database of free triadic conversations in various noise conditions. It contains movement data and separately available speech and background noise signals. This a priori knowledge enables a simulation of the conversations in a virtual acoustic environment. The resulting, separated signals at each receiver are then processed with implementations of scene-aware DRC (individual processing of foreground and background signals) and conventional DRC.
Results: The SNR benefit of processed signals is analyzed in five noise conditions: four of them are diffuse background noise (62 dB and 78 dB SPL babble noise and speech shaped noise) and one is speech dominated noise (a competing triadic conversation). Scene-aware DRC shows the highest SNR benefit across all conditions.
Conclusion: The higher SNR benefit in diffuse and speech dominated background noise indicates that performance benefits of scene-aware DRC hold up in complex communication scenarios. This result will be investigated with perceptual tests in future work.
