Speaker
Description
Recent advances in auditory modelling and deep neural networks have enabled the development of new hearing loss compensation methods. In these approaches, the parameters of a neural network are optimized so that the auditory representation of its output for a hearing-impaired listener is as close as possible to the normal-hearing auditory representation of clean speech. Previous studies have demonstrated the potential benefits of such systems, for example Leer et al. (2025) and Drakopulos (2025), using physiological auditory models during training.In this work, we propose a hearing loss compensation method based on the psychoacoustical loudness perception model AUDMOD (Bramsløw, 2004). The method was evaluated both objectively, using HASPI, and in a listening test with ten hearing-impaired participants, , using the Danish Sentence Test (DAST) and an AB preference test. The tested algorithms included different combinations of the proposed hearing loss compensation method and neural-network-based noise reduction, and their performance was compared with NAL-NL2.The resulting individual output signals were analyzed using several additional approaches, including spectral and temporal insertion-gain analysis and the assessment of modulation-frequency preservation. The findings reveal discrepancies between HASPI predictions and speech intelligibility measured in the listening tests.