14th Speech in Noise Workshop, 12-13 January 2023, Split, Croatia 14th Speech in Noise Workshop, 12-13 January 2023, Split, Croatia

P24Session 2 (Friday 13 January 2023, 09:00-11:00)
Towards predicting individual differences in the speech perception of hearing-impaired listeners

Helia Relaño-Iborra
Hearing Systems, Technical University of Denmark, Denmark

Johannes Zaar
Eriksholm Research Centre, Denmark

Torsten Dau
Hearing Systems, Technical University of Denmark, Denmark

The speech-based computational auditory signal processing and perception model [sCASP; Relaño-Iborra et al., (2019), J. Acoust. Soc. Am., 146(5), 3306–3317] was shown to successfully account for speech intelligibility measured in normal-hearing listeners in a wide range of listening conditions, including speech degradations as well as non-linear speech enhancement algorithms. The model consists on a chain of auditory-processing stages, including outer- and middle-ear filtering, a non-linear frequency-selective basilar membrane model, inner hair cell transduction and envelope modulation frequency filtering. The model’s decision metric is based on the cross-correlation between the clean and the degraded (or enhanced) speech representations in the modulation envelope domain. In this study, we evaluated the sCASP model as a predictor of speech intelligibility data measured in hearing-impaired (HI) listeners in a speech in noise task. The model was individually parametrized for each HI listener, based on their pure tone audiogram as well as estimates of cochlear compression and outer- and inner-hair cell loss. We analyzed the model’s predictive power in terms of its average performance across listener groups as well as by its ability to predict individual listeners’ speech reception thresholds (SRTs). The predictions obtained with sCASP accounted for the general decrease in performance observed for HI-listeners as compared to results from normal-hearing listeners. Furthermore, the model correctly predicted effects of different masker types in the SRT. Overall, the model accounted well for the trends observed at the group level, whereas reasonable correlations between the measured and predicted performance across the individual listeners were only found for a subset of the data. These promising results warrant further investigations into the sources underlying the across-listener variance and their potential integration in the model’s processing chain such that it can account for the performance of individual listeners.

Last modified 2023-01-06 23:41:06