Modelling the Effects of Hearing Aid Algorithms on Speech and Speaker Intelligibility as Perceived by Listeners with Simulated Sensorineural Hearing Impairment

The hearing-aid (HA) is the primary device used to improve sound perception for the sensorineural hearing-impaired (SHI) listener. However, despite the improvement of HA over the years, users continue to complain about their performance, especially in noisy environments; and the long time it takes to fit HAs. Testing and fitting HAs require evaluation by human subjects, which can take days or weeks. This study proposes replacing listeners with a computational model (CM) during the initial testing of HA algorithms or the fitting of HA. The aim is for the CM to predict listeners' speech and speaker intelligibility judgement and help reduce the time for testing and fitting HAs. The CM consists of a signal processing front end configured with the SHI listener's audiogram to mimic auditory threshold shifts and loudness recruitment of the HA wearer. The backend consists of a microscopic intelligibility model, that uses statistical speech models and knowledge of the background noise to make specific predictions about the words the listener perceived. Experiments were conducted using Grid corpus utterances, mixed with speech shaped noise (SSN) at a signal to noise ratio (SNR) of 0 dB. NAL-RP and a spectral envelope decimation (SED) frequency lowering HA algorithms were evaluated. Twenty normal-hearing listeners participated in the experiment. Two high-frequency hearing loss profiles were simulated, gently sloping, hi1, and steeply sloping, hi2. The CM's and listeners' performances were compared using different intelligibility measures. Statistically significant outcomes were obtained for average keyword intelligibility across HA algorithms for hi1 and hi2 simulated HI. Statistically significant results were also obtained for some individual keywords, and speakers' intelligibility using different HA algorithms and hi1 and hi2. The model also made a modest prediction of the NAL-RP and SED algorithms' performance in some of the tests. These results indicate that the CM has room for improvement; however, its success in some of the tests indicates the methodology used in its implementation may inform the development of a commercial application to conduct rapid testing and fitting of HAs