Modelling the effect of hearing impairment for binaural speech intelligibility in noise

Understanding speech amongst others is a challenging and daily situation occurring in public transport, food court, pub... A listener is able to segregate the target speech from the competing sources, so-called maskers, when they are spatially separated thanks to auditory mechanisms that use interaural level and time differences (ILD and ITD) of the signals. These mechanisms are well known as better-ear listening and binaural unmasking, respectively. However, their benefits are degraded when a listener suffers from hearing loss so that speech intelligibility can be greatly reduced. The motivation for this PhD is to develop a binaural speech intelligibility model that can account for the effects related to hearing loss in complex, realistic scenarios, including reverberation and competing masking sources. This model would help to better understand what aspects of hearing loss impact speech intelligibility.

The project has been structured in three main studies. The first study was about the optimization of a binaural speech intelligibility model for normal-hearing (NH) listeners applied to datasets testing auditory better-ear listening and binaural unmasking in isolation and combination, thus, hearing impairment was discarded in a first place. The model inputs are the speech and masker signals alone. They are decomposed per time frame and frequency band. Then, better-ear listening is modelled by taking the higher signal-to-noise ratio (SNR) between ears and the binaural unmasking advantage is estimated using a formula previously developed in the literature. The values are integrated across frequency, averaged across time and summed to provide a binaural ratio that can be compared to speech intelligibility threshold using a scaling method. It must be seen as a deep introduction of the original model that was further developed in the following studies to take into account hearing loss.

The second study investigated the influence of hearing loss on better-ear listening. The main outcome was the design of an internal noise level that can account for the effect of individual reduced audibility. The internal noise level is spectrally shaped on the listener’s audiogram and relies on the external stimulus level. In this version, the better-ear SNR is computed using the higher level between internal noise and external masker and the binaural unmasking advantage is computed only if the external signal levels are above the internal noise level at each ear. The model implementation was tested using three experiments from the literature involving NH and hearing-impaired (HI) listeners and varying the spatial separation of the sources, the masker type as well as the sensation level.

The last study highlights the contribution of binaural unmasking to speech intelligibility for NH and HI listeners. For this purpose, a dataset was collected during the project including ITD sensitivity and speech intelligibility measurements. Speech intelligibility was measured varying the presentation level of the signals, the masker type, the difference in ITD between speech and masker as well as the reverberation of the room. A new model version was developed making the two jitters of the formula computing the binaural unmasking advantage in the model dependent of the external stimulus level. This allowed to account for the effect of low presentation level on binaural unmasking for NH listeners.