Predicting Speech Intelligibility Based on Across-Frequency Contrast in Simulated Auditory-Nerve Fluctuations

Abstract

Summary The present study proposes a modeling approach for predicting speech intelligibility for normal-hearing (NH) and hearing-impaired (HI) listeners in conditions of stationary and fluctuating interferers. The model combines a non-linear model of the auditory periphery with a decision process that is based on the contrast across characteristic frequency (CF) after modulation analysis in the range of the fundamental frequency of speech. Specifically the short-term across-CF correlation between noisy speech and noise alone is assumed to be inversely related to speech intelligibility. The model provided highly accurate predictions for NH listeners as well as largely plausible effects in response to changes in presentation level. Furthermore, the model could account for some of the main features in the HI data solely by adapting the peripheral model using a simplistic interpretation of the listeners’ hearing thresholds. The model’s predictive power may be substantially improved by refining the interpretation of the HI listeners’ profiles and the model may thus p rovide a valuable basis for quantitatively modeling effects of outer hair-cell and inner hair-cell loss on speech intelligibility.

A speech-based computational auditory signal processing and perception model

“…Furthermore, the level analysis showed that sCASP can account for the effects of stimulation level on SI, thus reflecting processing limitations of the system both at low presentation levels (reflecting audibility limitations) and at high presentation levels (reflecting loss of frequency resolution and increased upwards spread of masking). Similar realistic SI predictions at different presentation levels were shown by the models of Scheidiger et al (2018) and Zilany and Bruce (2007). Both approaches included non-linear stages in their auditory preprocessing, indicating that the active mechanisms reflected in the sCASP framework might be responsible for its level-dependent behavior.…”

“…The proposed sCASP model performs a crosscorrelation of the modulation-spectrograms of the template and target in its backend by means of an analysis of the relationship across frequency bands including both the time and audio-frequency axes. This approach is similar to the ones proposed in Carney (2018) and Scheidiger et al (2018). Previously, Chabot-Leclerc et al (2014) argued that in the condition of phase-jitter distortion the data could only be accounted for by the sEPSM framework when adding an explicit analysis across audio-frequencies, inspired by the work of Elhilali et al (2003).…”

“…The internal representations of the pre-processed signals contain the dimensions time, audio frequency, and modulation frequency. Recent studies showed that the relative differences in the internal representation across the different auditory channels provide critical information related to the intelligibility of speech signals (e.g., Elhilali et al, 2003;Chabot-Leclerc et al, 2014;Carney, 2018;Scheidiger et al, 2018). Thus, the sCASP model's backend was designed such that it assumes independent processing across modulation channels but analyzes the contribution of all auditory channels simultaneously for each of the modulation frequencies.…”

“…Yet another approach would be to make assumptions about the sources of a given hearing impairment and to simulate the effects of such deficits on the internal auditory representation of the signals. For example, Scheidiger et al (2018) proposed a model that employs an auditory nerve (AN) simulation (Carney, 1993;Zilany et al, 2014) followed by a bandpass modulation filter centered at a frequency (125 Hz) close to the fundamental frequency (F 0 ) of the target speech. The model was shown to account for speech reception thresholds (SRTs; i.e., the signal level at which 50% of speech units are correctly understood by the listener) obtained in NH listeners for different interferer types.…”

“…In the present study, a conceptually similar approach to the one proposed by Scheidiger et al (2018) is presented. This approach also utilizes a detailed non-linear auditory model to predict SI but considers clean speech as the reference instead of noise (alone), such that the model is not limited to conditions where noise is present.…”