Do models hear the noise? Supra-threshold components of hearing loss in speech recognition models
* Presenting author
Hearing-impaired listeners show decreased speech recognition performance, and particularly complain about communication difficulties. Recently, the effect of impaired hearing on speech recognition thresholds (SRTs) was examined with the German matrix test in a stationary noise for 315 ears. Two domains of assumed linear relationship between SRTs and the pure-tone average (PTA) were identified; listening in noise, and (effectively) listening in quiet. Here, the individual SRTs were predicted based on the audiogram with the framework for auditory discrimination experiments (FADE) and the speech intelligibility index (SII). Overall, the predictions with FADE were more accurate than with the SII, with root-mean-square errors (RMSE) of 5.6dB and 6.8dB, respectively. The RMSE with the SII was highest for steep hearing losses (RMSE=22.6dB), where FADE performed better (RMSE=6.7dB). In the listening-in-noise domain, FADE underestimates the linear relationship of SRTs and PTAs, while the SII overestimates them. While FADE assumes no supra-threshold deficits, the SII evidently assumes even a larger amount than the empirical data imply. These results suggest that supra-threshold deficits should be considered separately from the hearing threshold and on the average explain only a fraction of the effect of impaired hearing on SRTs. Individually, however, supra-threshold deficits appear to play an important role.