@inproceedings {Schädler1726_2016,
	year = {2016},
	author = {Schädler, Marc René and Warzybok, Anna and Brand, Thomas and Kollmeier, Birger},
	title = {Simulating Speech Recognition of Listeners with Impaired Hearing using an Automatic Speech Recognition System: Modelling Suprathreshold Deficiencies with Level Uncertainty},
	booktitle = {Assoc. Res. Otolaryng. MidWinter Meeting (ARO)},
	URL = {http://c.ymcdn.com/sites/www.aro.org/resource/resmgr/Abstract_Archives/UPDATED_2016_ARO_Abstract_Bo.pdf},
	abstract = {Introduction
Which information is needed to complement the pure tone
audiogram in order to accurately predict speech reception
thresholds (SRTs) in noise for listeners with impaired hearing?
To answer this question, a framework which is able to
simulate speech recognition in noise of listeners with normal
hearing and accurately predict reference-free SRTs in several
noise conditions (Schädler et al., in press) was enhanced to
perform simulations for listeners with impaired hearing.
Method
Reference-free, i.e. not depending on any reference
measurement, SRTs were predicted by simulating the
German matrix sentence test (Wagener et al. 1999) with an
individually adapted automatic speech recognition (ASR)
system. In a first experiment, the effect of the absolute hearing
thresholds, determined from 200 individual audiograms, was
implemented into the feature extraction stage of the ASR
system. The recognition results were compared to empirical
data and speech intelligibility index (SII) based predictions
form the literature. For a second experiment, in extension
to the pure tone audiogram based simulations, a suprathreshold
deficiency, modelled by a level uncertainty uL in the
feature extraction stage of the ASR system, is postulated.
The values for the individual supra-threshold deficiency uL
were determined from the respective other noise condition by
choosing uL to match the empirical and predicted SRTs.
Results
The results show that the loss of sensitivity described by
the individual pure tone threshold explains only 40% of the
empirically observed variance in a stationary and 50% in a
fluctuating noise condition. Taking into account pure tone
threshold and individual supra-threshold hearing deficiency,
67% and 76% of the empirically observed variance can be
explained in the stationary and the fluctuating noise condition,
respectively.
Conclusions
The study concludes that the individual audiogram
describes the impaired hearing incompletely and should be
complemented with a description of individual supra-threshold
hearing deficiencies to improve the prediction accuracy. The
accurate simulation of human speech recognition, using suboptimal
signal processing in order to model impaired hearing,
could enable simulations under aided conditions. The
suitability of this approach for aided performance prediction
should be assessed in future work.
Funding
This work was supported by the Deutsche Forschungsgemeinschaft
SFB/TRR 31 “The active auditory system” and
the Cluster of Excellence Grant “Hearing4all”.}
}