Predicting Speech Perception in Older Listeners with Sensorineural Hearing Loss Using Automatic Speech Recognition

The objective of this study was to provide proof of concept that the speech intelligibility in quiet of unaided older hearing-impaired (OHI) listeners can be predicted by automatic speech recognition (ASR). Twenty-four OHI listeners completed three speech-identification tasks using speech materials of varying linguistic complexity and predictability (i.e., logatoms, words, and sentences). An ASR system was first trained on different speech materials and then used to recognize the same speech stimuli presented to the listeners but processed to mimic some of the perceptual consequences of age-related hearing loss experienced by each of the listeners: the elevation of hearing thresholds (by linear filtering), the loss of frequency selectivity (by spectrally smearing), and loudness recruitment (by raising the amplitude envelope to a power). Independently of the size of the lexicon used in the ASR system, strong to very strong correlations were observed between human and machine intelligibility scores. However, large root-mean-square errors (RMSEs) were observed for all conditions. The simulation of frequency selectivity loss had a negative impact on the strength of the correlation and the RMSE. Highest correlations and smallest RMSEs were found for logatoms, suggesting that the prediction system reflects mostly the functioning of the peripheral part of the auditory system. In the case of sentences, the prediction of human intelligibility was significantly improved by taking into account cognitive performance. This study demonstrates for the first time that ASR, even when trained on intact independent speech material, can be used to estimate trends in speech intelligibility of OHI listeners.