A model predicting the effect of hearing impairment on binaural speech intelligibility in noise

An update to the preliminary model proposed by Lavandier et al. [Acta Acust. united Ac. 104, 909-913 (2018)] will be presented that allows predicting binaural speech intelligibility in noise for normal hearing (NH) and hearing impaired (HI) listeners as a function of the listener’s audiogram and the noise level.

The model inputs are the masker and target signals as well as the audiogram at the listener’s ears. An internal noise is used to model hearing impairment and considered as another masker by the model. The model applies a short-time frequency analysis of the signals, to predict the binaural masking level difference (BMLD) and the signal-to-noise ratio (SNR) at the better ear. If the target or masker level is lower than the internal noise level at a given ear, the BMLD is set to 0. The maximum between the masker and internal noise levels is considered to compute the SNR at the better-ear. The BMLD and better-ear SNR are summed, integrated across frequency with a SII-weighting and averaged across time to obtain a binaural ratio. The relative difference between binaural ratios are compared to speech reception threshold differences measured in listening tests.

Three experiments were considered here, which used headphones to simulate an anechoic room, enrolled NH and HI listeners, and involved 2 vocoded speech maskers either collocated with the target in front of the listener or spatially separated. The separated conditions were artificially designed: the left masker was played only on the left channel of the headphones and the right masker only on the right channel. In the first and third experiments, the NH maskers were played at 60 dB SPL and a linear amplification was applied on the HI stimuli depending on their hearing loss; while in the second experiment, masker and target were filtered in order to equalize the audibility of the stimuli across listeners, then they were played at four different sensation levels. The first experiment also involved two unmodulated speech-shaped noises. In the third experiment, three types of separated conditions were tested: the first was similar to the one used in the other experiments, the second used realistic binaural cues and the third involved stimuli with interaural time differences removed. The proposed binaural model was able to predict the experimental data satisfactorily in all experimental conditions for NH and HI listeners.