Binaural audio processing to improve speech in noise intelligibility: Design and experimental evaluation with normal-hearing listeners and hearing-aid users

Abstract

[EN] Understanding speech in noisy environments remains challenging for individuals with hearing loss, particularly in listening scenarios involving multiple competing sound sources. While modern hearing aids incorporate directional microphones and noise reduction methods, their effectiveness is limited by assumptions about the acoustic scene or by the degradation of spatial cues in users of bilateral devices. Current approaches, such as adaptive beamforming or machine learning-based source separation, rely on complex estimations of the location or the characteristics of the target sound. Furthermore, they often require multiple microphones per device or synchronized bilateral processing. These constraints limit their applicability and leave many users with suboptimal performance in multi-talker environments. In this thesis, we present a binaural audio processing method designed to improve speech intelligibility in noisy environments by attenuating contralateral sounds. The method consists of subtracting a weighted version of the contralateral signal from the ipsilateral signal. It operates in the frequency domain, with subtraction weights based on generic head-related transfer functions, and without requiring prior knowledge of the target or noise properties or their location. The method is first evaluated using objective metrics in simulated listening scenarios. Results demonstrate consistent improvements in the signal-to-noise ratio and the short-term objective intelligibility, particularly for spatially separated target and noise sources. These findings confirm the theoretical premise that contralateral sound attenuation can improve speech intelligibility without the need for complex spatial analysis or source tracking. The method is then evaluated experimentally on individuals with normal hearing as well as in users of bilateral hearing aids. In normal-hearing listeners, the method improved intelligibility in unilateral listening but was less effective in bilateral listening due to its altering spatial cues. Hearing-aid users exhibited substantial improvements in both unilateral and bilateral listening, underscoring the method’s potential to compensate for degraded binaural processing. Overall, the findings support the effectiveness and practical relevance of the proposed method. Its simplicity and low computational cost make it a promising candidate for implementation in hearing technologies. The thesis contributes a novel sound processing method for speech enhancement that addresses real-world auditory challenges faced by individuals with and without hearing loss.