A model-based evaluation of speech perception in noise for electro-acoustic listeners
Due to recent advances in surgical techniques, there is a possibility to preserve existing residual acoustic hearing in many cochlear implant (CI) candidates even after implantation. This group of listeners who receives both electric and acoustic stimulation in the same ear is called electro-acoustic (EA) listeners. For these listeners, many clinical studies reported a benefit in EA-listening condition in comparison with electric-only or acoustic-only listening conditions (termed EA-benefit) for speech intelligibility in noise.
The goal of this study is to introduce a physiologically inspired auditory model of speech intelligibility that can predict the speech-in-noise perception of EA-listeners. In addition to the assessment of the effect of different physiological factors (residual acoustic hearing, spatial spread of electric field) on speech intelligibility, the model could help to investigate the underlying mechanism for EA-benefit.
Two different auditory models are used to simulate combined electrically/acoustically stimulated auditory nerve (AN) spikes. The auditory model of Fredelake and Hohmann (2012) simulates the AN spikes in response to electric stimulation. As this model mimics the CI signal processing strategy with a constant electric pulse rate, only the information about the envelope of the speech signal is extracted. The Meddis (2006) model is used to produce AN spikes in response to acoustic stimulation, mostly restricted to low frequencies for EA-listeners. This acoustically stimulated AN spiking pattern contains vital information about the fundamental frequency and in some cases up to first formant, which may not be available to conventional CI users in this form. The AN spiking patterns are further processed by a central auditory processing stage (Fredelake and Hohmann, 2012). This results in an internal representation (IR) of the stimuli. Speech reception thresholds in stationary noise were predicted by simulating the German matrix sentence test with an adapted automatic speech recognition system (Schädler, IJA, in press), using the individual or concatenated IRs as front-ends.
The model predicts an EA-benefit of up to 3 dB, a result which is in line with clinical studies. Changing the upper frequency boundary of residual acoustic hearing in the model showed that even a very restricted frequency range (which contains only information about the fundamental frequency) can still result in EA-benefit. Increasing the electric field spatial spread resulted in higher SRTs both in electric-only and electro-acoustic listening conditions. Overall the model could reproduce most of the typical SRTs and EA-benefits reported by clinical studies.
Supported by DFG cluster of excellence EXC 1077/1 “Hearing4All”